# Best of Agent Harnesses

> Hand-curated, ranked list of 164 AI agent harnesses — the runtimes that close the loop between a stateless model and the outside world. 12 categories, a 4-tier adoption-surface rating (simplicity ↔ capability), capability tags, a license signal, and one concrete example link per project. Stars captured 2026-09-09.

Maintained at https://github.com/RyanAlberts/best-of-Agent-Harnesses (CC-BY-SA-4.0).
Structured data: https://raw.githubusercontent.com/RyanAlberts/best-of-Agent-Harnesses/main/harnesses.json
Tiers, least to most adoption surface: super simple → mostly simple → slightly complex → complex.

## Pick by use case

- I want a turnkey coding agent today: opencode (https://github.com/anomalyco/opencode), Cline (https://github.com/cline/cline), Codex (https://github.com/openai/codex), Gemini CLI (https://github.com/google-gemini/gemini-cli), OpenHands (https://github.com/OpenHands/OpenHands), crush (https://github.com/charmbracelet/crush)
- I want an always-on personal agent that lives in my chat apps: OpenClaw (https://github.com/openclaw/openclaw), Hermes (https://github.com/NousResearch/hermes-agent), Khoj (https://github.com/khoj-ai/khoj), Agent Zero (https://github.com/agent0ai/agent-zero), OpenHarness (HKUDS) (https://github.com/HKUDS/OpenHarness)
- I want to extend Claude Code, Codex, or OpenCode with skills and slash commands: Anthropic Skills (https://github.com/anthropics/skills), wshobson/agents (https://github.com/wshobson/agents), superpowers (https://github.com/obra/superpowers), GStack (https://github.com/garrytan/gstack), pmstack (https://github.com/RyanAlberts/pmstack)
- I want to build my own coding harness from scratch: Claude Agent SDK (https://github.com/anthropics/claude-agent-sdk-python), Google ADK (https://github.com/google/adk-python), AutoHarness (https://github.com/aiming-lab/AutoHarness), SWE-agent (https://github.com/SWE-agent/SWE-agent), RepoMaster (https://github.com/QuantaAlpha/RepoMaster), claw-code-agent (https://github.com/HarnessLab/claw-code-agent)
- I want a drop-in memory layer for agents: Mem0 (https://github.com/mem0ai/mem0), Graphiti (Zep) (https://github.com/getzep/graphiti), claude-mem (https://github.com/thedotmack/claude-mem), agentlog (https://github.com/RyanAlberts/agentlog), letta (https://github.com/letta-ai/letta)
- I want to plug hundreds to thousands of tools without context bloat: MCP-Zero (https://github.com/xfey/MCP-Zero), ToolGen (https://github.com/Reason-Wang/ToolGen), ToolRAG (https://github.com/antl3x/ToolRAG), langgraph-bigtool (https://github.com/langchain-ai/langgraph-bigtool)
- I want multi-agent orchestration: openai-agents-python (https://github.com/openai/openai-agents-python), crewAI (https://github.com/crewAIInc/crewAI), autogen (https://github.com/microsoft/autogen), Microsoft Agent Framework (https://github.com/microsoft/agent-framework), PraisonAI (https://github.com/MervinPraison/PraisonAI), agent-squad (https://github.com/2FastLabs/agent-squad)
- I want a general LLM app framework: langgraph (https://github.com/langchain-ai/langgraph), langchain (https://github.com/langchain-ai/langchain), llama-index (https://github.com/run-llama/llama_index), pydantic-ai (https://github.com/pydantic/pydantic-ai), agno (https://github.com/agno-agi/agno)
- I want low-code / visual workflows: langflow (https://github.com/langflow-ai/langflow), Flowise (https://github.com/FlowiseAI/Flowise), Dify (https://github.com/langgenius/dify), n8n (https://github.com/n8n-io/n8n)
- I want browser-using agents: browser-use (https://github.com/browser-use/browser-use), Stagehand (https://github.com/browserbase/stagehand), WebVoyager (https://github.com/MinorJerry/WebVoyager), puppeteer-real-browser-mcp (https://github.com/withLinda/puppeteer-real-browser-mcp-server)
- I want sandboxed code execution for agent-generated code: E2B (https://github.com/e2b-dev/E2B), Agent Sandbox (https://github.com/kubernetes-sigs/agent-sandbox), Daytona (https://github.com/daytonaio/daytona), smolagents (https://github.com/huggingface/smolagents), OpenHands (https://github.com/OpenHands/OpenHands)
- I want to evaluate or benchmark agents: SWE-bench (https://github.com/SWE-bench/SWE-bench), Terminal-Bench (https://github.com/harbor-framework/terminal-bench), AgencyBench (https://github.com/GAIR-NLP/AgencyBench), inspect_ai (https://github.com/UKGovernmentBEIS/inspect_ai), WebArena (https://github.com/web-arena-x/webarena), VitaBench (https://github.com/meituan-longcat/vitabench)
- I want a deep research / autonomous research agent: deepagents (https://github.com/langchain-ai/deepagents), gpt-researcher (https://github.com/assafelovic/gpt-researcher), openagents (https://github.com/OpenAgentsInc/openagents)
- I want a provider-agnostic LLM pipe (not a framework): LiteLLM (https://github.com/BerriAI/litellm), vercel/ai (https://github.com/vercel/ai)

## Decision guides

- Agent evals: SWE-bench vs inspect_ai vs Terminal-Bench: https://ryanalberts.github.io/best-of-Agent-Harnesses/compare/agent-eval-harnesses/ — You changed your agent: new model, new prompt, new tools. Did it get better or worse? An eval is how you answer that with a number instead of a feeling. The word covers two different products, and knowing which one you need is most of the decision. A **benchmark** is a fixed public exam with a leade
- Browser agents: browser-use vs Stagehand vs Playwright MCP vs chrome-devtools-mcp: https://ryanalberts.github.io/best-of-Agent-Harnesses/compare/browser-agents/ — "Browser agent" covers three different kinds of product, and most bad picks here come from comparing across the lanes instead of within one. **Agent libraries** own the whole job: you hand them a goal in plain language ("find the cheapest flight, fill the form") and they decide every click. browser-
- Browser infrastructure for agents: Browserbase vs Steel vs Hyperbrowser: https://ryanalberts.github.io/best-of-Agent-Harnesses/compare/browser-infrastructure/ — Agent libraries like browser-use and Stagehand decide what to click; something still has to run the browsers they click in. At small scale that's Chrome on your own machine. At production scale it becomes its own operations problem: hundreds of concurrent sessions, sites that block automation, CAPTC
- Claude Code skill packs: superpowers vs GStack vs get-shit-done vs Anthropic Skills: https://ryanalberts.github.io/best-of-Agent-Harnesses/compare/claude-code-skill-packs/ — A skill is a folder of instructions (a SKILL.md file, plus any scripts it needs) that a coding agent loads only when the task matches, instead of carrying every instruction all the time. A skill pack is a curated bundle of them: someone else's working habits for your agent, installed as files. Nothi
- Eval and observability platforms: Langfuse vs LangSmith vs Braintrust vs Phoenix: https://ryanalberts.github.io/best-of-Agent-Harnesses/compare/eval-platforms/ — Benchmarks tell you how a model ranks; your production agent still fails in ways no public exam covers. An eval and observability platform is where teams watch what their agent actually did (tracing: recording every step, tool call, and token of a run) and score it continuously (evals: checks that r
- How to pick a harness: https://ryanalberts.github.io/best-of-Agent-Harnesses/compare/how-to-pick-a-harness/ — This is the decision guide for [best-of-Agent-Harnesses](../README.md), a curated, ranked list of the runtimes that turn an AI model into a working agent: the software that decides what the model's thinking is allowed to touch. Six questions, in order. Each one eliminates most of the list; by the en
- How to test-drive a harness: https://ryanalberts.github.io/best-of-Agent-Harnesses/compare/how-to-test-drive-a-harness/ — Spec sheets cannot answer "which harness should I use," because an agent's performance is a property of the *pairing* between harness and model, not of either alone: the same model passes [46% of tasks in one harness and 80% in another](https://www.mindstudio.ai/blog/agent-harness-scaffolding-matter
- Agent memory layers: Mem0 vs Zep vs Letta vs claude-mem: https://ryanalberts.github.io/best-of-Agent-Harnesses/compare/memory-layers/ — Agents forget. A model keeps nothing between sessions, so anything your agent should still know tomorrow (who the user is, what was decided, what failed last time) has to live in a memory system outside the model. "Add memory to my agent" then hides genuinely different products, and picking by star 
- Multi-agent orchestration: OpenAI Agents SDK vs CrewAI vs AutoGen vs Agent Framework vs LangGraph: https://ryanalberts.github.io/best-of-Agent-Harnesses/compare/multi-agent-orchestration/ — Orchestration is the layer that coordinates several AI agents working on one job: who acts next, what they share, and what happens when a step fails. The five frameworks here answer that with genuinely different architectures, and the pick is expensive to get wrong because the coordination model sha
- OpenClaw vs Hermes: the always-on personal-agent debate: https://ryanalberts.github.io/best-of-Agent-Harnesses/compare/openclaw-vs-hermes/ — An always-on personal agent is a program that runs all day on your own machine, talks to you through the chat apps you already use (WhatsApp, Telegram, Slack, Discord), and keeps acting between your messages: watching inboxes, running scheduled jobs, remembering what you told it last month. OpenClaw
- Context files for agents: AGENTS.md vs CLAUDE.md vs skills vs MCP tool search: https://ryanalberts.github.io/best-of-Agent-Harnesses/compare/progressive-disclosure/ — A model has a context window: a fixed amount of text it can consider at once. Everything competes for that space: your instructions, the definitions of every tool the agent could call, and the output of every tool it already called. Context bloat is what happens when the "might need it" pile crowds 
- Agent sandboxing: what it is and how to pick: https://ryanalberts.github.io/best-of-Agent-Harnesses/compare/sandboxed-code-execution/ — An AI agent does not just suggest code. It runs code, opens web pages, and edits files on a real computer. Agent sandboxing means making that computer a disposable one, so that when the agent makes a mistake, or a malicious web page talks it into one, the damage lands in a box you can throw away ins
- Terminal coding agents: opencode vs Codex vs Gemini CLI vs crush vs goose: https://ryanalberts.github.io/best-of-Agent-Harnesses/compare/terminal-coding-agents/ — The most-asked pick in this list: *"I want a turnkey coding agent in my terminal today."* A terminal coding agent is a program you run in your shell that takes a plain-language request, then works in a loop: the model proposes an action (edit this file, run this command, search the repo), the tool e

## FAQ

### What is the best agent harness if I want a turnkey coding agent today?
Top picks: opencode, Cline, Codex. See the “Coding agent products (IDEs, CLIs, full suites)” category for the full ranked list.

### What is the best agent harness if I want an always-on personal agent that lives in my chat apps?
Top picks: OpenClaw, Hermes, Khoj. See the “Personal agent runtimes” category for the full ranked list.

### What is the best agent harness if I want to extend Claude Code, Codex, or OpenCode with skills and slash commands?
Top picks: Anthropic Skills, wshobson/agents, superpowers. See the “Coding harness configs and SDKs” category for the full ranked list.

### What is the best agent harness if I want to build my own coding harness from scratch?
Top picks: Claude Agent SDK, Google ADK, AutoHarness. See the “Coding harness configs and SDKs” category for the full ranked list.

### What is the best agent harness if I want a drop-in memory layer for agents?
Top picks: Mem0, Graphiti (Zep), claude-mem. See the “Plugins, MCPs, CLI tools” category for the full ranked list.

### What is the best agent harness if I want to plug hundreds to thousands of tools without context bloat?
Top picks: MCP-Zero, ToolGen, ToolRAG. See the “Progressive disclosure harnesses” category for the full ranked list.

### What is the best agent harness if I want multi-agent orchestration?
Top picks: openai-agents-python, crewAI, autogen. See the “Multi-agent and orchestration” category for the full ranked list.

### What is the best agent harness if I want a general LLM app framework?
Top picks: langgraph, langchain, llama-index. See the “Frameworks” category for the full ranked list.

### What is the best agent harness if I want low-code / visual workflows?
Top picks: langflow, Flowise, Dify. See the “Frameworks” category for the full ranked list.

### What is the best agent harness if I want browser-using agents?
Top picks: browser-use, Stagehand, WebVoyager. See the “Plugins, MCPs, CLI tools” category for the full ranked list.

### What is the best agent harness if I want sandboxed code execution for agent-generated code?
Top picks: E2B, Agent Sandbox, Daytona. See the “Libraries and SDKs” category for the full ranked list.

### What is the best agent harness if I want to evaluate or benchmark agents?
Top picks: SWE-bench, Terminal-Bench, AgencyBench. See the “Evaluation and benchmarking harnesses” category for the full ranked list.

### What is the best agent harness if I want a deep research / autonomous research agent?
Top picks: deepagents, gpt-researcher, openagents. See the “Research and task-specific harnesses” category for the full ranked list.

### What is the best agent harness if I want a provider-agnostic LLM pipe (not a framework)?
Top picks: LiteLLM, vercel/ai. See the “Libraries and SDKs” category for the full ranked list.

### Which agent harnesses can run unattended (headless)?
Harnesses designed for unattended runs, batches, and fleets: opencode, OpenHands, goose, Symphony, SWE-agent, Claude Agent SDK, RepoMaster, OpenClaw.

### Which agent harnesses survive a crash mid-task (durable)?
Harnesses whose execution state persists across restarts: langgraph-bigtool, n8n, langgraph, mastra, letta, deepagents, pydantic-ai, Cloudflare Agents.

### How many of these agent harnesses are open source?
121 of 164 carry a standard open-source license; the rest are source-available or unclear, and flagged per row.

### What is an agent harness?
The runtime that turns a model into an agent: it decides what the model's reasoning is allowed to touch, and supplies the orchestration, tool wiring, memory, error recovery, and guardrails around per-turn inference.

### How is this list ranked?
By relevance to harness concerns (environment, orchestration, lifecycle, guardrails) and by GitHub stars (captured 2026-09-09); each project also carries an adoption-surface tier and autonomy/recovery scores.

### How can an AI agent use this list directly?
Three machine-readable surfaces: harnesses.json (structured), llms.txt (one file), and an MCP server (uvx agent-harnesses-mcp) exposing recommend, compare, pick_harness, and search_harnesses.

## Progressive disclosure harnesses (8 projects)

Formats, runtimes, and patterns that reveal context, tools, or instructions in layers—index first, details on demand—to control tokens and improve agent focus (the "map, not encyclopedia" principle).

- [Headroom](https://github.com/headroomlabs-ai/headroom) — ⭐71k, mostly simple, autonomy: n/a, recovery: n/a, open-source: Compresses tool outputs, logs, files, and RAG chunks with content-aware compressors before they reach the model—claimed 20% fewer tokens for coding agents and 60–95% fewer for JSON, same answers. Ships as a library, HTTP proxy, or MCP server, so it drops in front of whatever harness you already run. [mcp, rag]
- [awesome-cursorrules](https://github.com/PatrickJS/awesome-cursorrules) — ⭐40.7k, super simple, autonomy: n/a, recovery: n/a, open-source: Curated .cursorrules and skills that leverage Cursor's index-then-load model; the canonical collection for rules-as-progressive-disclosure in the IDE. [ide]
- [agents.md](https://github.com/agentsmd/agents.md) — ⭐24.2k, super simple, autonomy: n/a, recovery: n/a, open-source: Open format for repo-scoped agent briefings; nested AGENTS.md files scope instructions per directory, so agents get a map of what exists and load only what's relevant. Read by 20+ tools including Codex, Cursor, and Copilot. [ide, typescript]
- [context-mode](https://github.com/mksglu/context-mode) — ⭐21.6k, mostly simple, autonomy: n/a, recovery: n/a, restricted (Elastic-2.0): Context-window optimization layer that sandboxes tool output before it reaches the model (claimed 98% reduction) and persists session memory across 17 agent platforms via MCP and hooks—progressive disclosure applied to tool results, not just instructions. [mcp, memory, sandbox]
- [langgraph-bigtool](https://github.com/langchain-ai/langgraph-bigtool) — ⭐557, slightly complex, autonomy: bounded, recovery: durable, open-source: Build LangGraph agents with large tool sets; retrieval and on-demand tool loading so agents scale beyond context without stuffing every schema upfront. [tool-discovery, python]
- [MCP-Zero](https://github.com/xfey/MCP-Zero) — ⭐513, complex, autonomy: bounded, recovery: none, open-source: Active tool discovery for autonomous agents: model requests tools by requirement; hierarchical semantic routing over 308 servers / 2,797 tools with ~98% token reduction (APIBank). [tool-discovery]
- [ToolGen](https://github.com/Reason-Wang/ToolGen) — ⭐184, complex, autonomy: n/a, recovery: n/a, unknown: ICLR 2025: unified tool retrieval and calling via generation; 47k+ tools without context stuffing—retrieval and invocation in one generative step. [tool-discovery, python]
- [ToolRAG](https://github.com/antl3x/ToolRAG) — ⭐33, mostly simple, autonomy: n/a, recovery: n/a, open-source: Semantic tool retrieval for LLMs; serves only the tools the user query demands (MCP-compatible), unlimited tool sets with zero context penalty. [mcp, tool-discovery]

## Coding agent products (IDEs, CLIs, full suites) (23 projects)

Turnkey coding agents you install and run: IDE extensions, terminal CLIs, Dockerized workspaces. Each entry notes which part is the harness (the agent loop, tool wiring, approval model) versus the UI shell (VS Code extension, TUI, browser client).

- [opencode](https://github.com/anomalyco/opencode) — ⭐206k, slightly complex, autonomy: headless, recovery: resumable, open-source: Open-source terminal coding agent (formerly `sst/opencode`; transferred to anomalyco). The **harness** is a multi-provider tool-call loop (Claude, OpenAI, Gemini, local) with strong plugin and MCP support; the TUI is the shell. 100% OSS, very actively shipped. [mcp, provider-agnostic, cli, tui, typescript]
- [Codex](https://github.com/openai/codex) — ⭐123k, slightly complex, autonomy: bounded, recovery: resumable, open-source: OpenAI's terminal coding agent. The **harness** is the sandboxed tool-call loop with multi-provider support; the CLI is the shell. Reference implementation for "official CLI that ships code." [sandbox, provider-agnostic, cli]
- [Gemini CLI](https://github.com/google-gemini/gemini-cli) — ⭐107k, slightly complex, autonomy: bounded, recovery: resumable, open-source: Google's first-party terminal agent for Gemini. The **harness** is the plugin/MCP tool-call loop; the terminal is the shell—Google's parallel to Claude Code / Codex, not just an API. [mcp, cli, typescript]
- [pi](https://github.com/earendil-works/pi) — ⭐103k, slightly complex, autonomy: bounded, recovery: resumable, unknown: The upstream AI agent toolkit behind this list's oh-my-pi fork: a unified multi-provider LLM API, agent loop, and TUI shell providing the **harness** that oh-my-pi's Rust rewrite builds on. [provider-agnostic, tui, rust]
- [OpenHands](https://github.com/OpenHands/OpenHands) — ⭐87k, complex, autonomy: headless, recovery: resumable, restricted ((multi-license)): Dockerized software-engineering agent. The **harness** is the bash/editor/browser toolset with micro-agents and event-stream session bridging; Docker is the sandbox. Main OSS choice for teams self-hosting autonomous repo work. [memory, browser, sandbox, python]
- [Open Interpreter](https://github.com/openinterpreter/openinterpreter) — ⭐68.3k, mostly simple, autonomy: bounded, recovery: resumable, open-source: Lightweight terminal coding agent oriented to open models (DeepSeek, Kimi, Qwen). The **harness** is a code-execution loop — the model writes code, the harness executes it with confirmation gates; the CLI is the shell. The original "let the LLM run code on my machine" project, reborn for open weights. [cli, python]
- [Cline](https://github.com/cline/cline) — ⭐67.7k, slightly complex, autonomy: step-gated, recovery: resumable, open-source: VS Code extension whose **harness** is a plan-then-act loop with per-step human approval and cost transparency; the VS Code integration is the UI shell. Open-source counterweight to Cursor. [ide, typescript]
- [goose](https://github.com/aaif-goose/goose) — ⭐54.1k, slightly complex, autonomy: headless, recovery: resumable, open-source: Block-originated Rust agent, now stewarded by the Linux Foundation's Agentic AI Foundation (`aaif-goose/goose`). The **harness** is the MCP/ACP extension model with recipes and provider choice; there's no fixed UI slot—you bolt it into whatever shell you use. [mcp, rust]
- [DeepSeek-Reasonix](https://github.com/esengine/DeepSeek-Reasonix) — ⭐35.5k, slightly complex, autonomy: n/a, recovery: n/a, unknown: DeepSeek-native terminal coding agent. The **harness** is engineered around prefix-cache stability for long-running sessions; the TUI is the shell. [memory, cli, tui, typescript]
- [oh-my-pi](https://github.com/can1357/oh-my-pi) — ⭐30.3k, slightly complex, autonomy: bounded, recovery: resumable, open-source: Terminal coding agent (fork of Pi) that wires the IDE into the **harness**: hash-anchored edits, a 32-tool loop tuned per-model, LSP rename/references/diagnostics on every write, a real DAP debugger (lldb/dlv/debugpy), long-lived Python + Bun execution kernels that call back into the agent's tools, browser control, and 40+ providers (Claude/OpenAI/Gemini/local). ~55k-line Rust core. [browser, provider-agnostic, cli, ide, rust, python]
- [vibe-kanban](https://github.com/BloopAI/vibe-kanban) — ⭐28k, slightly complex, autonomy: n/a, recovery: n/a, unknown: Kanban-style fleet manager for running Claude Code, Codex, or any coding agent across many tasks at once. The **harness** contribution is the task-queue/review layer on top of whichever agent executes; not an agent loop itself.
- [crush](https://github.com/charmbracelet/crush) — ⭐28k, slightly complex, autonomy: bounded, recovery: resumable, restricted (FSL-1.1-MIT): Charm's terminal coding agent (Charm's fork of the original OpenCode). The **harness** is the tool-calling loop with session persistence; the Bubble Tea TUI is the shell. [memory, cli, tui]
- [qwen-code](https://github.com/QwenLM/qwen-code) — ⭐27.7k, slightly complex, autonomy: bounded, recovery: resumable, unknown: Alibaba's official terminal coding agent, forked from Gemini CLI's agent loop and retuned for Qwen models. The **harness** is the same sandboxed tool-call loop as its upstream; the terminal is the shell. [sandbox, cli, typescript]
- [Kilo Code](https://github.com/Kilo-Org/kilocode) — ⭐27.2k, slightly complex, autonomy: step-gated, recovery: resumable, unknown: VS Code extension and CLI in the Cline/Roo-Code lineage — a natural pick now that Roo-Code is archived upstream. The **harness** is an approval-gated autonomous-mode loop with a provider/tool marketplace; the IDE is the shell. [mcp, cli, ide, typescript]
- [Symphony](https://github.com/openai/symphony) — ⭐27.1k, complex, autonomy: headless, recovery: resumable, unknown: OpenAI's harness for fanning a task out into many isolated, autonomous coding-agent implementation runs and surfacing the ones that pass, so a team manages outcomes instead of supervising each session. [sandbox]
- [Roo Code](https://github.com/RooCodeInc/Roo-Code) — ⭐24.3k, slightly complex, autonomy: step-gated, recovery: resumable, open-source: VS Code/Cursor extension in the Cline lineage. The **harness** is the approval-gated agent with custom modes and a strong MCP story; the IDE is the UI. Popular community fork when you want that workflow without the upstream extension. [mcp, workflow, ide, typescript]
- [jcode](https://github.com/1jehuang/jcode) — ⭐19.4k, slightly complex, autonomy: n/a, recovery: n/a, unknown: Rust terminal coding agent pitched as the most RAM-efficient **harness** in its class; MCP support, multi-provider (Claude/OpenAI). [mcp, memory, provider-agnostic, cli, rust]
- [eigent](https://github.com/eigent-ai/eigent) — ⭐15.2k, complex, autonomy: n/a, recovery: n/a, unknown: Open-source desktop **harness** positioned as a local, free alternative to Claude Cowork and Codex: multi-agent workspace orchestration in a self-hosted app rather than a hosted product. [multi-agent, local]
- [cc-haha](https://github.com/NanmiCoder/cc-haha) — ⭐14.3k, complex, autonomy: n/a, recovery: n/a, unknown: Local-first desktop workspace **harness** for Claude Code and other agents: multi-agent sessions, Git worktrees, code diffs, a skill marketplace, and chat-app access (WeChat, Telegram, WhatsApp). [memory, multi-agent, typescript]
- [claw-code-agent](https://github.com/HarnessLab/claw-code-agent) — ⭐544, slightly complex, autonomy: checkpoint-gated, recovery: none, unknown: Python reimplementation of the Claude Code agent architecture with zero external dependencies; interactive chat, streaming, plugin runtime, nested agent delegation, cost tracking, MCP transport—portable harness without the Rust/TS toolchain. [mcp, rust, python, typescript]
- [Proliferate](https://github.com/proliferate-ai/proliferate) — ⭐464, complex, autonomy: bounded, recovery: resumable, open-source: Open-source AI IDE for Claude Code, Codex, OpenCode, and more. The **harness** contribution is the workspace/session orchestration layer: run multiple coding agents in parallel, locally or in the cloud, with isolated workspaces, reusable workflows, and shared team context. [multi-agent, sandbox, ide, typescript]
- [AgentBox](https://github.com/madarco/agentbox) — ⭐389, slightly complex, autonomy: n/a, recovery: n/a, open-source: Runs multiple coding agents in parallel, each in its own sandboxed VM, locally or in the cloud, from one command. The **harness** contribution is the VM-per-agent isolation and fleet fan-out layer; whichever agent runs inside owns the loop. [sandbox, typescript]
- [YYLO](https://github.com/yylo-dev/yylo) — ⭐58, slightly complex, autonomy: n/a, recovery: n/a, open-source: Command-line orchestrator for coding agents (drives Pi and Codex subagents): every task gets a dedicated branch/worktree and a typed lifecycle — read-only preflight, risk-scaled merge-queue review, release-readiness gates — with receipt-backed changes. The **harness** contribution is the git-native task/merge orchestration layer on top of whichever agent executes; not an agent loop itself. [multi-agent, typescript]

## Coding harness configs and SDKs (17 projects)

Skill packs, slash-command libraries, meta-prompting frameworks, and official SDKs that give you the harness (the agent loop, planning, memory, hooks) without bundling a specific IDE or CLI shell.

- [superpowers](https://github.com/obra/superpowers) — ⭐284k, complex, autonomy: n/a, recovery: n/a, open-source: Performance-oriented harness pack for Claude Code and 13 other harnesses (Codex, Cursor, OpenCode, Gemini CLI, more): skills, instincts, memory, security, research-first workflows. Treats harness engineering itself as the performance lever. [memory, cli, ide]
- [Anthropic Skills](https://github.com/anthropics/skills) — ⭐175k, mostly simple, autonomy: n/a, recovery: n/a, restricted (Anthropic terms): Anthropic's official Agent Skills repository: SKILL.md-based folders (instructions, scripts, resources) Claude dynamically loads on Claude Code, Claude.ai, and the API. The reference for progressive-disclosure skill packs in 2026.
- [GStack](https://github.com/garrytan/gstack) — ⭐132k, slightly complex, autonomy: n/a, recovery: n/a, open-source: Garry Tan's Claude Code skill stack: 23 slash-command modes (CEO/eng/design review, QA, ship, browse, retro, …) that structure one assistant as a virtual engineering team. Daily driver while running YC. [typescript]
- [addyosmani/agent-skills](https://github.com/addyosmani/agent-skills) — ⭐93.2k, mostly simple, autonomy: n/a, recovery: n/a, open-source: Addy Osmani's production-grade skill pack: 24 engineering skills and 4 specialist agent personas that encode senior-dev workflows (spec through deploy) across 70+ coding agents including Claude Code, Cursor, and Copilot. The **harness** contribution is the skill/workflow layer, not a new agent loop. [workflow, ide]
- [awesome-claude-code](https://github.com/hesreallyhim/awesome-claude-code) — ⭐53.8k, super simple, autonomy: n/a, recovery: n/a, unknown: Large community-curated index of Claude Code skills, slash commands, status lines, and plugins—resources for extending the harness, not a harness itself, but the most-followed catalog of the genre.
- [wshobson/agents](https://github.com/wshobson/agents) — ⭐39.5k, super simple, autonomy: n/a, recovery: n/a, open-source: Cross-harness marketplace of drop-in subagents and skills for Claude Code, Codex CLI, Cursor, OpenCode, and Copilot; specialized, production-ready agent definitions you install rather than hand-write. [multi-agent, cli, ide]
- [planning-with-files](https://github.com/OthmanAdi/planning-with-files) — ⭐26.7k, mostly simple, autonomy: n/a, recovery: n/a, unknown: Skill for persistent, file-based planning across long-running coding-agent sessions: crash-proof markdown plans, session recovery after `/clear`/compaction, and a deterministic completion gate—Manus-style planning as a drop-in **harness** layer via the Agent Skills standard. [memory]
- [SWE-agent](https://github.com/SWE-agent/SWE-agent) — ⭐20.3k, slightly complex, autonomy: headless, recovery: resumable, open-source: LM-driven harness built for SWE-bench: edit state, command execution, and issue-focused loop—the reference agent stack next to the benchmark itself. [memory, evals, python]
- [get-shit-done](https://github.com/open-gsd/gsd-core) — ⭐9.3k, mostly simple, autonomy: bounded, recovery: resumable, open-source: Goal-backward planning and wave-based execution over fresh context windows; avoids context rot by design. Python/JS meta-prompting for Claude Code, OpenCode, Gemini CLI. [cli, python]
- [Claude Agent SDK](https://github.com/anthropics/claude-agent-sdk-python) — ⭐8.1k, complex, autonomy: headless, recovery: resumable, open-source: Official Anthropic SDK (Python + [TypeScript](https://github.com/anthropics/claude-agent-sdk-typescript), [demos](https://github.com/anthropics/claude-agent-sdk-demos), [quickstarts](https://github.com/anthropics/claude-quickstarts)): built-in tools, MCP, long-running coding agents with session bridging. [mcp, memory, python, typescript]
- [agents-cli](https://github.com/google/agents-cli) — ⭐5.8k, mostly simple, autonomy: n/a, recovery: n/a, unknown: Google's official CLI and skill pack that layers agent-creation, evaluation, and deployment skills on top of whatever coding assistant you already run, rather than shipping its own agent loop—the **harness** as a config/skills add-on, not a new runtime. [evals, cli]
- [skillhub](https://github.com/iflytek/skillhub) — ⭐5.1k, mostly simple, autonomy: n/a, recovery: n/a, unknown: iFlytek's self-hosted registry for publishing, versioning, and governing agent skill packages—the **harness** config layer treated as an enterprise artifact store rather than a CLI or IDE shell. [local, cli, ide]
- [Meta-Harness](https://github.com/stanford-iris-lab/meta-harness) — ⭐1.5k, slightly complex, autonomy: n/a, recovery: n/a, unknown: Reference implementation from the Meta-Harness paper: an academic testbed for harness-engineering research, not a product—useful as a citation-grade baseline rather than something you'd run in production.
- [RepoMaster](https://github.com/QuantaAlpha/RepoMaster) — ⭐545, slightly complex, autonomy: headless, recovery: none, unknown: Repo-scoped research harness: builds function-call and module-dependency graphs to explore only what's needed; large relative gains on MLE-bench and GitTaskBench with lower token use. [workflow, python]
- [AutoHarness](https://github.com/aiming-lab/AutoHarness) — ⭐372, super simple, autonomy: bounded, recovery: none, open-source: Lightweight governance harness: wraps any LLM client in ~2 lines for automated harness engineering—6–14 step pipeline, YAML constitution, risk-pattern matching, session persistence with cost tracking, multi-agent profiles. [memory, multi-agent, provider-agnostic, python]
- [LoopTroop](https://github.com/looptroop-ai/LoopTroop) — ⭐133, mostly simple, autonomy: bounded, recovery: retry, open-source: Config layer that chains LLM councils for planning, Ralph loops for iterative refinement, and OpenCode worktrees for shipping. The **harness** contribution is the council → loop → worktree pipeline; OpenCode underneath executes. [typescript]
- [pmstack](https://github.com/RyanAlberts/pmstack) — ⭐8, super simple, autonomy: n/a, recovery: n/a, open-source: Claude Code config for AI product managers: CLAUDE.md plus skills for competitive analysis, PRD-from-signal, metric frameworks, stakeholder briefs, and agent eval design. "GStack for PMs." [evals]

## Personal agent runtimes (11 projects)

Always-on, self-hosted agents you run as a daemon and talk to from chat apps: gateway runtimes, second brains, and self-improving assistants. The agent as a product you operate, not a library you build with.

- [OpenClaw](https://github.com/openclaw/openclaw) — ⭐389k, complex, autonomy: headless, recovery: resumable, open-source: Self-hosted, always-on personal agent (formerly Clawdbot/Moltbot): a gateway + event-loop runtime that treats messages, heartbeats, crons, and webhooks as one input queue, persists state to local files, and lives in your chat apps (WhatsApp, Telegram, Slack, Discord). 13,700+ community skills; the fastest-growing repo in GitHub history. [typescript, multi-agent]
- [Hermes](https://github.com/NousResearch/hermes-agent) — ⭐244k, slightly complex, autonomy: headless, recovery: resumable, open-source: Nous Research's self-improving agent: a learning loop turns experience into reusable skills, builds a persistent user model across sessions, and checkpoints state to disk with rollback; lean enough for a $5 VPS, driven from chat, and model-agnostic (Nous Portal, OpenRouter, OpenAI, or any endpoint). [memory, python, provider-agnostic]
- [AnythingLLM](https://github.com/Mintplex-Labs/anything-llm) — ⭐65.8k, complex, autonomy: headless, recovery: resumable, open-source: Self-hosted "AI second brain" **harness**: chat with your documents, run built-in agent skills (web search, code execution, browsing), and manage multi-user workspaces with a bundled vector DB, so there is no separate RAG stack to wire up. [rag, typescript]
- [nanobot](https://github.com/HKUDS/nanobot) — ⭐47.9k, mostly simple, autonomy: n/a, recovery: n/a, unknown: Ultra-lightweight, self-hosted personal agent framework: the **harness** is a Python daemon wiring tools, memory, and MCP into chat/webhook front ends (Telegram, Discord, web); minimal footprint alternative to heavier personal-runtime stacks. [mcp, memory, local, python]
- [CowAgent](https://github.com/zhayujie/CowAgent) — ⭐46.9k, slightly complex, autonomy: n/a, recovery: n/a, unknown: Self-hosted **harness** (formerly chatgpt-on-wechat) that plans tasks, runs tools/skills, and self-evolves via memory; multi-model, multi-channel (WeChat, Telegram, etc.), one-line install. [memory, python]
- [Khoj](https://github.com/khoj-ai/khoj) — ⭐37.2k, complex, autonomy: headless, recovery: resumable, open-source: Self-hostable "AI second brain": answers over your docs and the web, custom agents, scheduled automations, and multi-client reach (web, Obsidian, Emacs, WhatsApp). A personal-agent harness with retrieval at the core. [python]
- [Eliza](https://github.com/elizaOS/eliza) — ⭐19.3k, complex, autonomy: headless, recovery: resumable, open-source: Open "agentic operating system" (elizaOS): persistent multi-agent runtime with character files, a plugin ecosystem, and social/platform integrations — the harness behind a large share of autonomous social agents. [memory, multi-agent, typescript]
- [Agent Zero](https://github.com/agent0ai/agent-zero) — ⭐19.1k, slightly complex, autonomy: bounded, recovery: resumable, unknown: Organic, prompt-defined personal agent framework: hierarchical sub-agents, persistent memory, browser and code tools, and self-modifying behavior; runs in Docker with a web UI. [memory, multi-agent, browser, sandbox, python]
- [OpenHarness (HKUDS)](https://github.com/HKUDS/OpenHarness) — ⭐15.7k, complex, autonomy: bounded, recovery: resumable, open-source: Open agent harness with a built-in personal agent ("Ohmo") that runs across Feishu, Slack, Telegram, and Discord; core tool-use, skills, memory, multi-agent coordination with auto-compaction for multi-day sessions. [memory, multi-agent]
- [AIlice](https://github.com/myshell-ai/AIlice) — ⭐1.4k, slightly complex, autonomy: bounded, recovery: none, open-source: Fully autonomous general-purpose agent; one binary, Docker-ready, for when you want "set goal and walk away" without a framework. [sandbox, python]
- [Talon](https://github.com/dylanneve1/talon) — ⭐76, slightly complex, autonomy: headless, recovery: resumable, open-source: Multi-platform personal agent living in Telegram, Discord, Teams, and the terminal. The **harness** is a pluggable-backend loop (Claude, Kilo, OpenCode, Codex, OpenAI Agents) with full MCP tool access and persistent background agents (Goals, Heartbeat, Dream); the chat apps are shells. [mcp, memory, cli, typescript]

## Frameworks (26 projects)

General-purpose agent and LLM application frameworks (the app layer, not harnesses per se).

- [n8n](https://github.com/n8n-io/n8n) — ⭐204k, complex, autonomy: headless, recovery: durable, restricted (Fair-code): Fair-code workflow engine with 400+ nodes and native AI nodes; the self-hosted Zapier that actually does agents and LangChain. [workflow, local, typescript]
- [AutoGPT](https://github.com/Significant-Gravitas/AutoGPT) — ⭐187k, complex, autonomy: headless, recovery: resumable, restricted (Polyform-SU): The original autonomous loop: goal in, agent iterates with tools and memory; Forge is the dev framework, Benchmark the eval harness. [memory, evals, python]
- [Dify](https://github.com/langgenius/dify) — ⭐155k, complex, autonomy: headless, recovery: retry, restricted (Fair-code): One-stop LLM app platform: visual workflows, RAG pipeline, 50+ tools, model management; "ship from prototype to prod" in a single UI. [low-code, rag, python]
- [langflow](https://github.com/langflow-ai/langflow) — ⭐154k, complex, autonomy: headless, recovery: retry, open-source: Low-code UI to build and deploy LangChain/LangGraph flows; visual DAG editor and one-click run. [low-code, python]
- [langchain](https://github.com/langchain-ai/langchain) — ⭐146k, complex, autonomy: bounded, recovery: retry, open-source: Chains, tools, retrievers, and agents; the usual entry point for "add tools to an LLM" in Python/JS. [python]
- [browser-use](https://github.com/browser-use/browser-use) — ⭐114k, slightly complex, autonomy: bounded, recovery: retry, open-source: Python web-agent **harness**: natural-language goals become browser actions, driven directly over the Chrome DevTools Protocol (it dropped Playwright in August 2025). The biggest community in the browser-agent category. [browser, python]
- [Flowise](https://github.com/FlowiseAI/Flowise) — ⭐55.4k, complex, autonomy: headless, recovery: retry, restricted (Apache+CLA): Drag-and-drop LangChain UI; deploy flows without code. The low-code sibling to Langflow, with a different component and hosting story. [low-code, typescript]
- [llama-index](https://github.com/run-llama/llama_index) — ⭐52.1k, complex, autonomy: bounded, recovery: retry, open-source: Data-centric: indexing, RAG, and query engines; agent abstractions sit on top of your data pipelines. [rag, python]
- [agno](https://github.com/agno-agi/agno) — ⭐42.1k, complex, autonomy: bounded, recovery: resumable, open-source: Python agents with memory, knowledge bases, tools, and structured outputs; continues the PhiData-era product line under the Agno name—production apps, evals, and pipelines. [memory, evals, python]
- [langgraph](https://github.com/langchain-ai/langgraph) — ⭐41.3k, slightly complex, autonomy: headless, recovery: durable, open-source: State-machine graphs over LLM steps; checkpointing, human-in-the-loop, and durable execution so workflows survive restarts. [workflow, python]
- [semantic-kernel](https://github.com/microsoft/semantic-kernel) — ⭐28.5k, complex, autonomy: bounded, recovery: retry, open-source: Microsoft's plugin and planner layer for LLMs; C#, Python, Java; strong on enterprise auth and orchestration. [python]
- [mastra](https://github.com/mastra-ai/mastra) — ⭐27.9k, slightly complex, autonomy: bounded, recovery: durable, restricted (Elastic-2.0): TypeScript-first; agents, tools, and workflows with a single runtime and minimal boilerplate. [typed, typescript]
- [Haystack](https://github.com/deepset-ai/haystack) — ⭐26.5k, complex, autonomy: n/a, recovery: n/a, open-source: Open-source orchestration framework for context-engineered LLM apps: modular pipelines and agent workflows with explicit control over retrieval, routing, memory, and generation—closer to LangChain's territory than a coding-agent harness. [memory, rag, python]
- [letta](https://github.com/letta-ai/letta) — ⭐24.7k, mostly simple, autonomy: headless, recovery: durable, open-source: Python agent runtime with tool use and control flow; lean API; stateful agents with long-horizon memory. [memory, python]
- [Stagehand](https://github.com/browserbase/stagehand) — ⭐24.2k, slightly complex, autonomy: bounded, recovery: none, open-source: Browserbase's SDK for browser agents: natural-language actions (act, extract, observe) and deterministic Playwright code mix in one script, so agent flexibility and repeatable automation live in the same **harness**. [browser, typescript]
- [Google ADK](https://github.com/google/adk-python) — ⭐21.5k, complex, autonomy: headless, recovery: resumable, open-source: Google's official Agent Development Kit: code-first Python toolkit for building, evaluating, and deploying agents. Optimized for Gemini but model-agnostic; deploys to Cloud Run / Vertex AI; ships a dev UI with eval and a code-execution sandbox. [evals, sandbox, python]
- [rasa](https://github.com/RasaHQ/rasa) — ⭐21.3k, complex, autonomy: headless, recovery: resumable, open-source: Conversational AI stack (NLU, dialogue, actions); long-standing OSS choice for chat and voice bots. [voice, python]
- [botpress](https://github.com/botpress/botpress) — ⭐14.9k, complex, autonomy: headless, recovery: resumable, open-source: Visual bot builder and runtime; multi-channel, open-source alternative to commercial bot platforms. [low-code, typescript]
- [R2R](https://github.com/SciPhi-AI/R2R) — ⭐8k, complex, autonomy: headless, recovery: retry, open-source: RAG-first: hybrid search, knowledge graphs, multimodal; the framework for "production RAG" when you care more about retrieval than chat UI. [vision, rag, workflow, python]
- [agent-squad](https://github.com/2FastLabs/agent-squad) — ⭐7.8k, slightly complex, autonomy: bounded, recovery: resumable, open-source: AWS-originated orchestrator (now under 2FastLabs): intent classification, streaming, SupervisorAgent; "agent-as-tools" so one agent delegates to a squad. [multi-agent]
- [AgentVerse](https://github.com/OpenBMB/AgentVerse) — ⭐5.1k, complex, autonomy: headless, recovery: none, open-source: Task-solving and simulation envs for multi-LLM agents; deploy many agents in custom environments without building infra from scratch. [multi-agent, python]
- [youtu-agent](https://github.com/TencentCloudADP/youtu-agent) — ⭐4.6k, mostly simple, autonomy: bounded, recovery: retry, unknown: Tencent Cloud's agent framework: a minimal tool-calling **harness** designed to perform well with open-source models, positioned as a lighter alternative to heavier orchestration frameworks.
- [Bee Agent Framework](https://github.com/i-am-bee/beeai-framework) — ⭐3.4k, complex, autonomy: bounded, recovery: resumable, open-source: Python + TypeScript, LF AI–backed; MCP/ACP, workflows, Requirement Agent; the one that pushes "production multi-agent" without LangChain. [mcp, multi-agent, python, typescript]
- [AgentStack](https://github.com/agentstack-ai/AgentStack) — ⭐2.2k, slightly complex, autonomy: n/a, recovery: n/a, open-source: Scaffolds full agent projects; plugs in CrewAI, LangGraph, OpenAI Swarm, LlamaStack and wires AgentOps observability from day one.
- [AgentSilex](https://github.com/howl-anderson/agentsilex) — ⭐455, super simple, autonomy: bounded, recovery: none, open-source: ~300 lines of readable agent code on top of LiteLLM; the "I want to see the whole loop" option for learning or minimal production. [python]
- [SuperAgentX](https://github.com/superagentxai/superagentx) — ⭐204, mostly simple, autonomy: bounded, recovery: none, open-source: Lightweight multi-agent orchestrator with an AGI-angle; minimal surface, docs-first, for teams that want orchestration without the kitchen sink. [multi-agent, python]

## Multi-agent and orchestration (12 projects)

Harnesses and patterns for multi-agent coordination and handoffs.

- [MetaGPT](https://github.com/FoundationAgents/MetaGPT) — ⭐70.3k, complex, autonomy: headless, recovery: resumable, open-source: The "AI software company" multi-agent framework: role-played PM, architect, and engineer agents turn a one-line requirement into specs, designs, and code along an SOP assembly line. The landmark of the genre; development pace has slowed in 2026. [multi-agent, python]
- [autogen](https://github.com/microsoft/autogen) — ⭐60.9k, complex, autonomy: bounded, recovery: resumable, open-source: Conversable agents and group chats; code execution and human-in-the-loop; Microsoft origin, AG2 ecosystem. ⚠️ In maintenance mode since late 2025: no new features, and Microsoft directs new users to Agent Framework. [multi-agent, python]
- [crewAI](https://github.com/crewAIInc/crewAI) — ⭐58.3k, complex, autonomy: bounded, recovery: resumable, open-source: Role-based agents (roles, goals, backstories) in Crews; Flows add event-driven and hierarchical control for production. [python]
- [OpenManus](https://github.com/FoundationAgents/OpenManus) — ⭐58.3k, complex, autonomy: bounded, recovery: none, open-source: Open, invite-free general agent from the MetaGPT team: planning plus tool use over a multi-agent loop, aimed at reproducing Manus-style autonomous task completion on your own keys. [multi-agent, python]
- [ChatDev](https://github.com/OpenBMB/ChatDev) — ⭐34.2k, slightly complex, autonomy: headless, recovery: none, open-source: Multi-agent software-company simulation (CEO, CTO, programmer, tester) built on chat chains with communicative dehallucination; ChatDev 2.0 continues the line. MetaGPT's conversational sibling. [python]
- [openai-agents-python](https://github.com/openai/openai-agents-python) — ⭐29.3k, mostly simple, autonomy: bounded, recovery: resumable, open-source: Handoffs, guardrails, and multi-LLM routing; minimal surface so you own the loop. [python]
- [Microsoft Agent Framework](https://github.com/microsoft/agent-framework) — ⭐13.4k, slightly complex, autonomy: bounded, recovery: resumable, open-source: Microsoft's convergence of AutoGen and Semantic Kernel: build, orchestrate, and deploy agents and multi-agent workflows in Python and .NET, with graph-based workflows and checkpointing — the designated successor harness for both lines. [multi-agent, workflow, python]
- [hive](https://github.com/aden-hive/hive) — ⭐11k, complex, autonomy: bounded, recovery: resumable, unknown: Self-hosted multi-agent **harness** aimed at production workloads: human-in-the-loop checkpoints and a self-improving agent loop, distinct from single-session coding-agent shells. [multi-agent, python]
- [omnigent](https://github.com/omnigent-ai/omnigent) — ⭐9.8k, complex, autonomy: n/a, recovery: n/a, unknown: Open-source meta-**harness**: orchestrates Claude Code, Codex, Cursor, Pi, and custom agents behind one policy/sandboxing layer so teams swap harnesses without rewriting workflows. [sandbox, ide, python]
- [PraisonAI](https://github.com/MervinPraison/PraisonAI) — ⭐9k, mostly simple, autonomy: bounded, recovery: none, open-source: Autonomous multi-agent teams with a single entry point; emphasis on minimal config. [multi-agent, python]
- [AG2](https://github.com/ag2ai/ag2) — ⭐4.9k, complex, autonomy: n/a, recovery: n/a, unknown: AG2 (formerly AutoGen): the community-governed continuation of the original AutoGen project after Microsoft's fork diverged—conversable multi-agent groups, code execution, and human-in-the-loop under an open-source AgentOS banner. Graduated off the radar this cycle. [multi-agent, python]
- [AgentRL](https://github.com/THUDM/AgentRL) — ⭐351, complex, autonomy: headless, recovery: resumable, open-source: Multitask, multiturn RL for LLM agents; Ray-based scaling, rollout/actor workers—for teams that want to train agents, not just run them. [training, python]

## Plugins, MCPs, CLI tools (19 projects)

IDE plugins, concrete MCP servers, and CLI tools that give agents tools and context.

- [MCP Servers](https://github.com/modelcontextprotocol/servers) — ⭐90.2k, mostly simple, autonomy: n/a, recovery: n/a, open-source: The official reference collection of Model Context Protocol servers (filesystem, git, fetch, memory, time, and more)—the canonical, vetted toolset agents connect to, and the pattern every other MCP server is measured against. [mcp, memory, typescript]
- [Context7](https://github.com/upstash/context7) — ⭐61.8k, super simple, autonomy: n/a, recovery: n/a, open-source: MCP server that injects up-to-date, version-specific library docs into an agent's context on demand; kills the stale-training-data hallucinations that plague codegen. [mcp, training, typescript]
- [chrome-devtools-mcp](https://github.com/ChromeDevTools/chrome-devtools-mcp) — ⭐51.4k, mostly simple, autonomy: n/a, recovery: n/a, open-source: Google's official Chrome DevTools MCP server: exposes console, network, and performance-trace inspection as tool calls agents can drive directly, instead of a human clicking through the DevTools panel. [mcp, browser, typescript]
- [aider](https://github.com/Aider-AI/aider) — ⭐48.9k, slightly complex, autonomy: checkpoint-gated, recovery: resumable, open-source: Git-aware CLI pair programmer; edits in-repo, supports multiple models and MCP so agents see version control and tools. [mcp, cli, python]
- [Playwright MCP](https://github.com/microsoft/playwright-mcp) — ⭐36.9k, mostly simple, autonomy: n/a, recovery: n/a, open-source: Playwright's official MCP server: structured browser control (navigate, click, fill, extract) via the accessibility tree rather than screenshots, so web tasks stay fast and deterministic. [mcp, vision, browser, typescript]
- [continue](https://github.com/continuedev/continue) — ⭐35.8k, complex, autonomy: checkpoint-gated, recovery: resumable, open-source: Open-source IDE extension (VS Code, JetBrains); in-editor completion and chat with local or API models. [ide, typescript]
- [github-mcp-server](https://github.com/github/github-mcp-server) — ⭐32.8k, slightly complex, autonomy: n/a, recovery: n/a, open-source: GitHub's official MCP server (Go): repos, issues, PRs, code search, Actions. Replaces the older community `cyanheads/github-mcp-server` as the canonical way to give agents GitHub access. [mcp]
- [MCP Python SDK](https://github.com/modelcontextprotocol/python-sdk) — ⭐24.3k, mostly simple, autonomy: n/a, recovery: n/a, open-source: Official SDK to build and consume MCP servers/clients in Python; stdio and SSE transports. [mcp, python]
- [MCP TypeScript SDK](https://github.com/modelcontextprotocol/typescript-sdk) — ⭐13.4k, mostly simple, autonomy: n/a, recovery: n/a, open-source: Official MCP implementation for Node/TS; reference for the protocol. [mcp, typescript]
- [MCP Inspector](https://github.com/modelcontextprotocol/inspector) — ⭐10.8k, super simple, autonomy: n/a, recovery: n/a, open-source: GUI to test and debug MCP servers; inspect tools, resources, and prompts. [mcp, typescript]
- [MCP Registry](https://github.com/modelcontextprotocol/registry) — ⭐7.2k, slightly complex, autonomy: n/a, recovery: n/a, open-source: Official, community-driven registry for MCP servers—the "app store" MCP clients use to discover servers. Maintained by Anthropic + ecosystem maintainers; v0.1 API frozen, production-grade. [mcp]
- [Agent Governance Toolkit](https://github.com/microsoft/agent-governance-toolkit) — ⭐6.2k, slightly complex, autonomy: n/a, recovery: n/a, open-source: Microsoft's policy-enforcement layer for autonomous agents: zero-trust identity, execution sandboxing, and OWASP Agentic Top-10 coverage sit in front of the agent loop as a **harness** security layer—the governance counterpart to Infisical's agent-vault. [sandbox, python]
- [mcp-context-forge](https://github.com/IBM/mcp-context-forge) — ⭐4.4k, complex, autonomy: n/a, recovery: n/a, unknown: IBM's official AI gateway/registry/proxy that sits in front of any MCP, A2A, or REST/gRPC API: unified endpoint, centralized discovery, guardrails, and plugin support—the enterprise front door for **harness** tool calling. Graduated off the radar this cycle. [mcp, python]
- [cocoindex-code](https://github.com/cocoindex-io/cocoindex-code) — ⭐2.7k, mostly simple, autonomy: n/a, recovery: n/a, unknown: Embedded, tree-sitter/AST-based code-search CLI and MCP server that gives coding agents fast semantic lookups over a repo instead of grepping or re-reading whole files into context. [mcp, cli]
- [agent-vault](https://github.com/Infisical/agent-vault) — ⭐2.2k, mostly simple, autonomy: n/a, recovery: n/a, unknown: Infisical's HTTP credential proxy that fronts secrets for Claude Code, OpenClaw, and other agent harnesses so the agent's tool calls never see raw credentials—a **harness** security layer, not an agent loop itself.
- [Docker MCP Gateway](https://github.com/docker/mcp-gateway) — ⭐1.6k, slightly complex, autonomy: n/a, recovery: n/a, open-source: Docker's official MCP CLI plugin / gateway; container-aware MCP tooling from Docker (replaces deprecated `docker/mcp-servers` path). [mcp, sandbox, cli]
- [puppeteer-real-browser-mcp](https://github.com/withLinda/puppeteer-real-browser-mcp-server) — ⭐26, mostly simple, autonomy: n/a, recovery: n/a, unknown: Puppeteer MCP with real-browser and anti-detection; for agents that need to drive sites that block headless. [mcp, browser, typescript]
- [Better-OpenCodeMCP](https://github.com/ajhcs/Better-OpenCodeMCP) — ⭐9, mostly simple, autonomy: n/a, recovery: n/a, open-source: MCP server for OpenCode/Crush: async task execution, model bridging (e.g. Claude→Gemini), process pooling. [mcp, typescript]
- [agentlog](https://github.com/RyanAlberts/agentlog) — ⭐1, super simple, autonomy: n/a, recovery: n/a, open-source: Persistent decision memory for any project: `remember`, `recall`, `reflect`. Single-file Python CLI that stores decisions as JSONL and uses Claude or Gemini to retrieve and synthesize patterns—Karpathy's LLM Wiki concept as a CLI. [memory, cli, python]

## Memory and state (5 projects)

Persistent memory layers that give agents recall across turns and sessions: knowledge graphs, vector stores, and session-capture tools that survive a restart. The state a harness needs but rarely ships with.

- [claude-mem](https://github.com/thedotmack/claude-mem) — ⭐93.6k, slightly complex, autonomy: n/a, recovery: n/a, open-source: Session-memory plugin for Claude Code, Codex, OpenClaw, Gemini, Copilot, and more: captures everything an agent does during a session, AI-compresses it, and injects the relevant context into future sessions. Session-to-session memory as a drop-in. [memory]
- [Mem0](https://github.com/mem0ai/mem0) — ⭐65k, slightly complex, autonomy: n/a, recovery: n/a, open-source: Universal memory layer for AI agents: stores user/org/session memory, retrieves on demand. Apache-2.0; the de-facto memory primitive paired with most harnesses in 2026. [memory, python]
- [Graphiti (Zep)](https://github.com/getzep/graphiti) — ⭐30.7k, slightly complex, autonomy: n/a, recovery: n/a, open-source: Zep's open-source memory engine: real-time temporal knowledge graphs that track how facts about users and entities change over time, so agents can answer "what was true when." The layer behind Zep's hosted memory platform. [memory, rag, workflow, python]
- [cognee](https://github.com/topoteretes/cognee) — ⭐30.6k, slightly complex, autonomy: n/a, recovery: n/a, open-source: Open-source memory layer for agents: an extract–cognify–load pipeline that turns your data into a queryable knowledge graph plus vector store, so agents recall facts and relationships across sessions instead of re-reading context. [memory, rag, workflow, python]
- [beads](https://github.com/gastownhall/beads) — ⭐27k, mostly simple, autonomy: n/a, recovery: n/a, unknown: Portable persistent-memory layer for coding agents: tracks decisions and task state outside the harness's own context window so it survives session resets and model swaps. [memory]

## Evaluation and benchmarking harnesses (19 projects)

Agentic eval systems, reasoning benchmarks, and open agent benchmarks.

- [Agent Lightning](https://github.com/microsoft/agent-lightning) — ⭐18k, complex, autonomy: headless, recovery: resumable, open-source: Microsoft's training-oriented harness: optimization loops for agent behavior—when you need to improve policies over rollouts, not only score a fixed prompt. [evals, training, python]
- [SWE-bench](https://github.com/SWE-bench/SWE-bench) — ⭐5.8k, slightly complex, autonomy: headless, recovery: resumable, open-source: LMs resolve real GitHub issues; Docker harness, instance IDs; standard for code-agent evals. [evals, sandbox, python]
- [AgentBench](https://github.com/THUDM/AgentBench) — ⭐3.7k, complex, autonomy: headless, recovery: none, open-source: ICLR'24 benchmark: agents across AlfWorld, DB, knowledge graphs, OS, webshop; Docker Compose, function-calling interface. [evals, sandbox, rag, workflow, python]
- [inspect_ai](https://github.com/UKGovernmentBEIS/inspect_ai) — ⭐2.7k, complex, autonomy: headless, recovery: resumable, open-source: Inspect AI core: composable eval tasks, sandboxes, scorers, and multi-model runs; the framework behind inspect_evals, not just the task bundle. [evals, sandbox, python]
- [WebArena](https://github.com/web-arena-x/webarena) — ⭐1.6k, complex, autonomy: headless, recovery: none, open-source: Realistic web env (e.g. e‑commerce, CMS, dev tools); 812 tasks; measures end-to-end web agent success. [python]
- [WebVoyager](https://github.com/MinorJerry/WebVoyager) — ⭐1.1k, slightly complex, autonomy: headless, recovery: none, open-source: End-to-end web agent with LMMs: screenshots + actions on real sites; benchmark on 15 sites, GPT-4V for automatic eval. [evals, vision]
- [agent-qa](https://github.com/vostride/agent-qa) — ⭐908, slightly complex, autonomy: headless, recovery: retry, restricted (FSL-1.1-ALv2): Self-improving QA **harness** for web and mobile apps: natural-language tests, memory-backed self-healing, dashboard/CLI, MCP and skills support, plus sandboxed hooks for production regression checks. [mcp, memory, sandbox, cli, typescript]
- [swe-smith](https://github.com/SWE-bench/SWE-smith) — ⭐764, slightly complex, autonomy: headless, recovery: none, open-source: Data generation for SWE agents; 50k+ instances across 128 repos; used for SWE-agent-LM training. [training, python]
- [ARC-AGI-2](https://github.com/arcprize/ARC-AGI-2) — ⭐738, super simple, autonomy: n/a, recovery: n/a, open-source: ARC Prize task set: grid-based abstraction/reasoning; public and private splits for generalization.
- [SWE-Gym](https://github.com/SWE-Gym/SWE-Gym) — ⭐733, slightly complex, autonomy: headless, recovery: none, open-source: Training and evaluation for SWE agents and verifiers (ICML 2025). [evals, training, python]
- [ClawBench](https://github.com/TIGER-AI-Lab/ClawBench) — ⭐694, complex, autonomy: headless, recovery: none, open-source: Open web-agent evaluation **harness**: runs selectable agents in isolated Docker containers across 153 live-site tasks (plus 130 in V2), intercepts irreversible requests, and records video, screenshots, HTTP traffic, actions, and agent messages for replayable scoring. [evals, vision, sandbox, python]
- [inspect_evals](https://github.com/UKGovernmentBEIS/inspect_evals) — ⭐664, slightly complex, autonomy: headless, recovery: resumable, open-source: UK AISI/Arcadia/Vector: GAIA and other evals in Inspect AI; level 1–3, sandboxed, tool-calling solvers. [evals, sandbox]
- [Terminal-Bench](https://github.com/harbor-framework/terminal-bench) — ⭐652, slightly complex, autonomy: headless, recovery: none, open-source: The terminal-task benchmark coding agents now cite next to SWE-bench: hard, containerized terminal tasks scored end to end. Terminal-Bench 2.0 runs on the harbor evaluation framework; the 1.0 tasks live on in the org's terminal-bench-1 repo. [evals, cli, python]
- [arc-agi-benchmarking](https://github.com/arcprize/arc-agi-benchmarking) — ⭐362, mostly simple, autonomy: headless, recovery: retry, open-source: Runner for ARC-AGI: multi-provider (OpenAI, Anthropic, Gemini, etc.), rate limits, retries, and scoring. [evals, provider-agnostic, python]
- [VitaBench](https://github.com/meituan-longcat/vitabench) — ⭐174, complex, autonomy: headless, recovery: none, open-source: ICLR'26: 66 tools, real-world apps (delivery, travel, retail); 100 cross-scenario + 300 single-scenario tasks; adopted by Qwen/Seed.
- [AgencyBench](https://github.com/GAIR-NLP/AgencyBench) — ⭐97, complex, autonomy: headless, recovery: none, open-source: Long-horizon agent benchmark: 32 scenarios, 138 tasks, ~1M tokens and ~90 tool calls; Docker sandbox and rubric-based + LLM judges. [evals, sandbox, python]
- [letta-evals](https://github.com/letta-ai/letta-evals) — ⭐82, mostly simple, autonomy: headless, recovery: none, open-source: Eval harness for stateful Letta agents; configurable suites and grading (LLM or rule-based) so you can measure what you ship. [memory, python]
- [SUPER](https://github.com/allenai/super-benchmark) — ⭐58, slightly complex, autonomy: headless, recovery: none, open-source: Agents that set up and run ML/NLP from GitHub repos; 45 expert problems, 152 masked tasks, 602 AutoGen tasks; Docker-based. [sandbox, python]
- [TRAIL](https://github.com/patronus-ai/trail-benchmark) — ⭐22, mostly simple, autonomy: n/a, recovery: n/a, open-source: Trace reasoning and agentic issue localization; 148 long-context traces, 841 errors, 20+ error types; Hugging Face dataset.

## Observability and eval-ops (4 projects)

Tracing, monitoring, and production evaluation for live agent runs: capture every step, tool call, and token, then score and debug in the loop. Distinct from the fixed-task benchmarks above—this is what you run against your own traffic.

- [Langfuse](https://github.com/langfuse/langfuse) — ⭐34.4k, slightly complex, autonomy: n/a, recovery: n/a, open-source: Open-source LLM engineering platform: full-trace observability, online and offline evals, prompt management, and cost metrics for agent runs in production—the monitoring layer most harnesses lack out of the box. [evals, typescript]
- [MLflow](https://github.com/mlflow/mlflow) — ⭐27.9k, complex, autonomy: n/a, recovery: n/a, open-source: Mature ML platform now covering GenAI: MLflow Tracing captures every agent step, tool call, and token, with built-in LLM evals and prompt versioning—observability for teams already standardized on MLflow. [evals, python]
- [Opik](https://github.com/comet-ml/opik) — ⭐21.9k, slightly complex, autonomy: n/a, recovery: n/a, open-source: Comet's open-source agent observability and evaluation platform: tracing, scoring, and experiment comparison with the whole core feature set free to self-host under Apache-2.0. [evals, python]
- [Arize Phoenix](https://github.com/Arize-ai/phoenix) — ⭐11.4k, slightly complex, autonomy: n/a, recovery: n/a, restricted (Elastic-2.0): Arize's source-available, local-first tracing and eval layer: run it on your laptop or your own infra, and graduate to the managed Arize AX platform only when you need it. [evals, python]

## Research and task-specific harnesses (5 projects)

Deep research, document QA, and domain-specific agent loops.

- [DeerFlow](https://github.com/bytedance/deer-flow) — ⭐82.1k, complex, autonomy: headless, recovery: resumable, open-source: ByteDance's long-horizon research **harness**: a LangGraph-based agent loop that researches, codes, and creates using sandboxes, memory, tools, and subagents behind a message gateway—an alternative to the reference gpt-researcher stack for multi-hour tasks. [memory, multi-agent, sandbox, python]
- [gpt-researcher](https://github.com/assafelovic/gpt-researcher) — ⭐29.4k, complex, autonomy: bounded, recovery: retry, open-source: Autonomous deep-research agent: web + local sources, citation-grounded reports, multi-agent and deep-research modes. The reference open-source research harness. [multi-agent, python]
- [AutoResearchClaw](https://github.com/aiming-lab/AutoResearchClaw) — ⭐14.4k, complex, autonomy: n/a, recovery: n/a, unknown: Fully autonomous, self-evolving research **harness**: multi-agent debate and citation verification carry a project from idea to a written paper without a human in the loop. [multi-agent]
- [MiroThinker](https://github.com/MiroMindAI/MiroThinker) — ⭐8.4k, slightly complex, autonomy: headless, recovery: retry, unknown: Deep-research **harness** tuned for long browsing-and-reasoning chains; benchmarked on BrowseComp, GAIA, and HLE by pairing a dedicated agent loop with its own MiroThinker models rather than bolting search onto a generic chat agent. [evals]
- [openagents](https://github.com/OpenAgentsInc/openagents) — ⭐446, complex, autonomy: headless, recovery: resumable, open-source: Platform for autonomous agents and autopilot-style workflows; decentralized/Nostr-oriented (Pylon runtime, actively shipped in 2026).

## Libraries and SDKs (15 projects)

Lightweight runtimes, tool loops, and provider-agnostic harness primitives.

- [Daytona](https://github.com/daytonaio/daytona) — ⭐71.7k, slightly complex, autonomy: n/a, recovery: n/a, open-source: Elastic dev environments for AI-generated code: workspaces, Git, previews—infra harness between "the model wrote a patch" and "it ran in a real machine." ⚠️ Public repo unmaintained since June 2026: core development moved to a private codebase (final open release v0.190.0, AGPL-3.0). [sandbox]
- [LiteLLM](https://github.com/BerriAI/litellm) — ⭐58.4k, mostly simple, autonomy: n/a, recovery: retry, open-source: One interface to 100+ LLMs; routing, caching, budgets. Not an agent framework—the pipe every agent framework uses. [provider-agnostic, python]
- [Composio](https://github.com/ComposioHQ/composio) — ⭐30.1k, complex, autonomy: n/a, recovery: n/a, open-source: 1,000+ toolkits with auth, tool search, and a sandboxed workbench—drop-in tool layer so agents stop reinventing OAuth + integrations. Python and TypeScript. [sandbox, tool-discovery, python, typescript]
- [smolagents](https://github.com/huggingface/smolagents) — ⭐29.3k, mostly simple, autonomy: bounded, recovery: none, open-source: Code-as-action agents: model outputs Python executed in sandbox (E2B, Modal, etc.); ~1k LOC core. [sandbox, python]
- [deepagents](https://github.com/langchain-ai/deepagents) — ⭐29.2k, slightly complex, autonomy: bounded, recovery: durable, open-source: LangChain's Python+TypeScript agent harness on top of LangGraph: planning tool, virtual filesystem, shell sandbox, sub-agent spawning—the "Claude Code-style" harness as a reusable library. [multi-agent, sandbox, python, typescript]
- [vercel/ai](https://github.com/vercel/ai) — ⭐26.7k, slightly complex, autonomy: bounded, recovery: retry, open-source: React and Node SDK for streaming, tool calls, and agent-style UIs; provider-agnostic. [provider-agnostic, typescript]
- [pydantic-ai](https://github.com/pydantic/pydantic-ai) — ⭐19.8k, slightly complex, autonomy: bounded, recovery: durable, open-source: Type-safe Python agents with Pydantic I/O; multi-provider, MCP, Logfire observability, and human-in-the-loop. [mcp, typed, provider-agnostic, python]
- [E2B](https://github.com/e2b-dev/E2B) — ⭐13.7k, slightly complex, autonomy: n/a, recovery: n/a, open-source: Firecracker sandboxes for executing agent-generated code; the hosted isolation layer many tool-calling demos use instead of running arbitrary LLM output on your laptop. [sandbox, python]
- [Steel](https://github.com/steel-dev/steel-browser) — ⭐7.6k, slightly complex, autonomy: n/a, recovery: n/a, open-source: Open-source browser API for agents: cloud or self-hosted Chrome sessions with stealth, residential proxies, CAPTCHA solving, and persistent profiles. The only open-source core in the hosted browser-infrastructure lane (Browserbase and Hyperbrowser are closed). [memory, browser, local]
- [strands-agents](https://github.com/strands-agents/harness-sdk) — ⭐7.2k, mostly simple, autonomy: bounded, recovery: resumable, open-source: Model-driven Python SDK; decorators for tools, native MCP, multi-agent; "minimal code" without sacrificing provider choice. [mcp, multi-agent, typed, python]
- [Cloudflare Agents](https://github.com/cloudflare/agents) — ⭐5.5k, slightly complex, autonomy: headless, recovery: durable, open-source: Persistent, stateful agents on Durable Objects: state, websockets, scheduling, and AI chat baked in. The serverless answer to "where does the agent live?" [memory, typescript]
- [Agent Sandbox](https://github.com/kubernetes-sigs/agent-sandbox) — ⭐3.8k, slightly complex, autonomy: n/a, recovery: n/a, open-source: Kubernetes-native sandbox primitive for agent runtimes: a Sandbox resource plus warm pools and claims for fast-start, isolated, stateful workloads. The self-hosted answer to hosted sandbox APIs, from the Kubernetes SIGs org. [memory, sandbox, local]
- [openai-agents-js](https://github.com/openai/openai-agents-js) — ⭐3.8k, slightly complex, autonomy: bounded, recovery: resumable, open-source: Official OpenAI Agents SDK for Node/TS: handoffs, guardrails, voice; the JS counterpart to openai-agents-python. [multi-agent, voice, typescript]
- [open-harness](https://github.com/MaxGfeller/open-harness) — ⭐610, slightly complex, autonomy: bounded, recovery: none, open-source: TypeScript Agent class on Vercel AI SDK; streaming events, filesystem/bash tools, MCP, and subagent delegation. [mcp, multi-agent, typescript]
- [Community-curated agent lists](https://github.com/brandonhimpfen/awesome-ai-agents) — ⭐15, super simple, autonomy: n/a, recovery: n/a, unknown: Broader directories: e.g. [brandonhimpfen/awesome-ai-agents](https://github.com/brandonhimpfen/awesome-ai-agents), [axioma-ai-labs/awesome-ai-agent-frameworks](https://github.com/axioma-ai-labs/awesome-ai-agent-frameworks), [mb-mal/awesome-ai-agents-frameworks](https://github.com/mb-mal/awesome-ai-agents-frameworks)—differ by scope and update cadence.