The open-source AI coding agent for VS Code — built local-first.
Chat with an agent that reads your workspace, edits files, runs commands, and searches your codebase semantically. Use your Claude / ChatGPT / Gemini subscription, any API key, or run completely offline with llama.cpp and Ollama.
Install · Local AI · Providers · Features · Contributing
🔌 100% Local AI
OpenCursor is designed to work without internet once set up:
- 🦙 llama.cpp built in — search Hugging Face for GGUF models, pick a quantization, download, and OpenCursor spawns and manages
llama-serverfor you. Full launch control: context size, GPU layers, flash attention, KV cache types, speculative decoding, vision (--mmproj), and more. - 🐋 Ollama — pull, manage, and chat with models from the Ollama library, zero config.
- 🧠 Local embeddings — semantic codebase search powered by an on-device ONNX MiniLM model. No embedding API, no key, no code leaving your machine.
- ✈️ Airplane-mode coding — local model + local index = a fully working AI agent, offline.
🔍 Semantic search, no cloud
Ask questions in plain language — "where do we refresh the auth token?" — and the agent automatically finds code by meaning, not keywords. The index builds automatically, updates incrementally, and is stored locally. Prefer a hosted embedding model? Point it at any OpenAI-compatible /embeddings endpoint instead.
🌐 Providers
- OAuth sign-in — connect your existing Claude Code, OpenAI Codex, or Google Antigravity account and use your subscription's models directly in VS Code.
- API keys — OpenAI, Anthropic, Gemini, OpenRouter presets out of the box.
- Custom providers — add any OpenAI-compatible or Anthropic-style endpoint (base URL + key). Run multiple providers at once; models are fetched live and mixable in the picker.
- Auto mode — a judge model routes each task to the best enabled model. Per-model reasoning effort, thinking mode, and context-size options.
⚡ What it does
| 💬 Agent chat | Multi-tab sidebar chat with @-mentions: files, folders, code, @Docs, git commits, branch diffs, terminals, rules, past chats |
| 🛠️ 25 tools | Read/write/edit, shell, grep/glob, semantic search, web search & fetch, notebooks, todos, subagents, MCP |
| ✅ Inline review | Per-hunk Keep / Undo CodeLenses on every agent edit — no git required |
| 🧭 Modes | Agent · Ask · Plan · Debug · Multitask (Ask/Plan are read-only) |
| 🛡️ Approval policy | Per-action allow/ask/review/deny with risk heuristics (rm -rf, sudo, .env, secrets…) and wildcard allow/deny lists |
| 🔗 MCP & hooks | Full MCP client, plus 11 lifecycle hooks (Cursor/Claude-Code compatible) and custom subagents |
| 🖼️ Images & PDFs | Paste screenshots, read images and PDFs directly into context |
| ⚙️ Extras | Ctrl+L adds selection to chat, personas, custom system prompt, token usage tracking, completion notifications |
Installation
- Install OpenCursor from the VS Code Marketplace (or grab the
.vsixfrom Releases). - Open the OpenCursor sidebar and pick a provider:
- Local: one-click llama.cpp install, or point at a running Ollama.
- Account: sign in with Claude Code / OpenAI Codex / Google Antigravity.
- API: paste a key for OpenAI, Anthropic, Gemini, OpenRouter, or any custom endpoint.
- Start chatting.
Ctrl+Lsends your selection to the chat.
Native runtime dependencies (ONNX runtime, image processing) are downloaded once on first activation with integrity checks — they're too heavy to ship in the VSIX.
Building from source
git clone https://github.com/PawanOsman/OpenCursor.git
cd OpenCursor
pnpm install
pnpm run compile # or: pnpm run watchPress F5 in VS Code to launch the Extension Development Host. Build a package with pnpm run vsix.
Contributing
Issues and PRs welcome — see issues.