UCAgent (UnityChip Verification Agent)

AI-powered automated UT verification agent based on large language models

中文介绍 | UCAgent Online Documentation

Introduction

UCAgent is an automated hardware verification AI agent based on large language models, focusing on Unit Test verification for chip design. It automatically analyzes hardware designs, generates test cases, executes verification tasks, and produces test reports through AI technology, thereby improving verification efficiency.

Key Features:

• Automated chip verification workflow
• Support for functional coverage and code coverage analysis
• Consistency assurance among documentation, code, and reports
• Deep collaboration with mainstream Code Agents (OpenHands, Copilot, Claude Code, Gemini-CLI, Qwen-Code, etc.) via MCP protocol
• Three intelligent interaction modes (standard, enhanced, advanced)

For more details, please refer to UCAgent Online Documentation

[!TIP] The slide, code and environment for UCAgent workshop presented at RISC-V Summit Europe 2026 are available at the official repository: https://github.com/XS-MLVP/tutorial-records

System Requirements

• Python 3.11+, if the latest version of Python encounters dependency issues, please downgrade
• Supported OS: Linux, macOS
• Memory: 4GB+ recommended
• Network: Access to AI model API (OpenAI compatible)
• picker: https://github.com/XS-MLVP/picker

Quick Start

1. Clone the Repository

git clone https://github.com/XS-MLVP/UCAgent.git
cd UCAgent

2. Install Dependencies

pip3 install -r requirements.txt

3. Install and Configure qwen

Please refer to https://qwenlm.github.io/qwen-code-docs/en/ to install qwen-code-cli, then configure the MCP Server as shown below.

Example ~/.qwen/settings.json:

{
    "mcpServers": {
           "unitytest": {
            "httpUrl": "http://localhost:5000/mcp",
            "timeout": 300000
        }
    }
}

Since running test cases may take a long time, it is recommended to set a larger timeout value, for example 300 seconds.

For other Code Agents, please refer to their documentation, e.g., claude code, opencode, copilot-cli, kilo-cli, iflow, etc.

4. Start Verification

Taking Adder in examples as an illustration.

4.1 Method 1: Automatically run qwen with specified backend (Recommended)

# Default backend is langchain,
#   requires configuration: OPENAI_API_BASE and other environment variables
# backend can be: langchian, claude, opencode, copilot, kilo, qwen, iflow, etc.
make mcp_Adder ARGS="--loop --backend=qwen"

For supported backends, please refer to the backend section in ucagent/setting.yaml.

4.2 Method 2: Manually run qwen (For unadapted CodeAgents)

（1）Start MCP-Server

make mcp_Adder  # workspace is set to output/workspace_Adder directory
# Calls the following commands:
#   picker export Adder/Adder.v --rw 1 --sname Adder --tdir output/workspace_Adder/ -c -w output/workspace_Adder/Adder/Adder.fst
#   ucagent output/workspace_Adder/ Adder -s -hm --tui --mcp-server-no-file-tools --no-embed-tools
# Browser Web UI mode:
#   ucagent output/workspace_Adder/ Adder -s -hm --web-console --mcp-server-no-file-tools --no-embed-tools
# Custom Web UI host/port/password (HTTP Basic Auth):
#   ucagent output/workspace_Adder/ Adder -s -hm --web-console 0.0.0.0:18000:secret --mcp-server-no-file-tools --no-embed-tools

The default MCP Server address is: http://127.0.0.1:5000/mcp

（2）Start qwen to execute task

cd output/workspace_Adder
qwen

After starting qwen as above, input the task prompt:

Please use the tool RoleInfo to get your role information and basic guidance, then complete the task. Use the tool ReadTextFile to read files. You need to perform file operations in the current working directory and should not go beyond this directory.

Note:

• Start the Code Agent in the working directory (e.g., output/workspace_Adder in the example above), otherwise file path mismatch issues may occur.
• If the DUT is complex and has peripheral component dependencies, you need to open the default skipped stages via ucagent interaction commands.

Tips:

• Write verification prompts according to task requirements
• When Code Agent stops midway, you can input: Continue, please use tool 'Complete' to determine if all tasks are finished

💡 More Usage Methods: Besides MCP collaboration mode, UCAgent also supports direct LLM integration, human-machine collaboration, and other modes. See Usage Documentation

5. How to Improve Verification Quality (Optional)

By default, UCAgent only enables the internal Python Checker for stage checking, which is heuristic. If you need verification quality improvement, you can enable LLM stage checking. If you need to reach "delivery level" quality, you further need to enable Human stage checking.

1. Enable LLM stage checking

2. Enable human stage checking

Default stage checking order: Python Checker -> LLM -> Human

Interact via Web Interface

UCAgent provides Master mode, based on which you can perform centralized Agent management, create tasks, view status, use online terminals, and other operations through the web interface.

Local Startup

1. Configure Environment Variables

# Edit a custom file to export environment variables required by ucagent, for example:
# export OPENAI_API_BASE=<your_openai_api_base>
# export OPENAI_API_KEY=<your_openai_api_key>
# export OPENAI_MODEL=<your_openai_model>
vim ~/.ucagent_env
# Then load the environment variables
source ~/.ucagent_env

2. Start UCAgent Master

make as_master_persist
# Or, if ucagent is installed, you can directly run ucagent to start master mode
ucagent --as-master-persist --as-master

Then visit http://localhost:8800 in your browser.

Docker Startup

docker run -it --rm \
  -e OPENAI_API_BASE=<your_openai_api_base> \
  -e OPENAI_API_KEY=<your_openai_api_key> \
  -e OPENAI_MODEL=<your_openai_model> \
  -p 8800:8800 \
  ghcr.io/xs-mlvp/ucagent:latest ucagent --as-master-persist --as-master

If ghcr.io is not accessible, you can directly replace it with mirror addresses such as ghcr.nju.edu.cn.

After successful startup, visit http://localhost:8800 in your browser.

Basic Operations

1. In the web interface, click the + button (or launch button) to create a new task.
2. In the Agent list, click the API button to connect to the control page of a specific Agent.
3. In the Agent control page, click the web terminal button to open the online terminal.
4. Start ucagent locally and connect to an existing Master service via the --master parameter.

More Features

1. Launch page: create workspace, upload/import files, parse modules, compile, and preview launch command.
2. Task page: filter, paginate, inspect task details/logs, and stop/delete managed tasks.
3. Enhanced Agent page: stage multi-select and bulk toggles (HM/Skip/LFail/LPass), plus stage artifact content/diff review.
4. Unified proxy access: Master proxies cmd/terminal/web-console paths for both task and agent entries.
5. Improved Web Terminal: multiple terminal sessions across different URLs.

📖 Detailed Operations: See TUI Usage Documentation

Basic Operations

TUI Shortcuts

• ctrl+up/down/left/right: Adjust layout (Console height / Mission panel width)
• ctrl+h/j/k/l: Vim-style layout adjustment (equivalent to ctrl+left/down/up/right)
• ctrl+c: Cancel running command; exit TUI if no command is running
• ctrl+t: Open theme picker
• f1: Show/hide keyboard shortcuts help panel
• shift+right: Clear console output
• shift+left: Clear input text
• tab: Command completion; press Tab repeatedly to cycle through candidates
• pageup/pagedown: Page through Console output
• esc: Exit scrolling/paging/help panel, or clear input

Stage Color Indicators

• White: Pending execution
• Red: Currently executing
• Green: Execution passed
• *:
  • Blue indicates LLM Fail checking is enabled for this stage, providing modification suggestions when stage check fails more than 3 times
  • Green indicates LLM Pass checking is enabled for this stage, verifying if stage task requirements are met upon completion
  • Red indicates this stage requires mandatory human inspection, AI can continue after entering command hmcheck_pass [msg]
• Yellow: Stage skipped

Common Interactive Commands

• q: Exit TUI (or exit UCAgent)
• tui: Enter TUI
• tab: Command completion
• tool_list: List all available tools
• help: View all command help
• loop [prompt]: Continue current task

📖 Detailed Operations: See TUI Usage Documentation

Frequently Asked Questions (FAQ)

Q: How to configure different AI models?

A: Modify the openai.model_name field in config.yaml, which supports any OpenAI-compatible API. See Configuration Documentation.

Q: What to do when errors occur during verification?

A: Use Ctrl+C to enter interactive mode, check current status with status, and use help to get debugging commands.

Q: MCP server cannot connect?

A: Check if the port is occupied, verify firewall settings, and you can specify a different port with --mcp-server-port.

Q: Why is there information from the last execution?

A: UCAgent by default looks for the .ucagent/ucagent_info.json file in the working directory to load previous execution information and continue. If you don't need history, delete this file or use the --no-history parameter to ignore loading history.

Q: How to run long-duration verification?

A: Please refer to CodeAgent's custom backend mode examples/CustomBackend/README.md.

Q: Can verification stages be customized?

A: Yes, see Customization Documentation.

Q: How to add custom tools?

A: Create a new tool class in the ucagent/tools/ directory, inherit from the UCTool base class, and load it with the --ex-tools parameter. See Tool List Documentation.

🔍 More Questions: Check the complete FAQ Documentation

Documentation Build and Preview (MkDocs)

The Makefile provides documentation-related helper targets (MkDocs + Material):

Usage Flow

First-time use (install dependencies):

make docs-install    # Install mkdocs and material theme dependencies

Daily development (preview documentation):

make docs-serve      # Start local server, visit http://127.0.0.1:8030
# Browser will auto-refresh after modifying docs

Local generation and viewing (build production version):

make docs-build      # Generate static website to docs/site directory
# Open docs/site/index.html in local browser
make docs-clean      # Clean build artifacts (optional)

Complete Workflow Example

# 1. Initial setup: Install dependencies
make docs-install

# 2. Development phase: Preview docs (can be repeated)
make docs-serve      # Visit http://127.0.0.1:8030 in browser
# ...edit documentation...
# Press Ctrl+C to stop service

# 3. Local generation: Build production version
make docs-build      # Generate docs/site directory
# Open docs/site/index.html in local browser

# 4. Cleanup (optional)
make docs-clean      # Delete docs/site directory

Notes

• Port and address are currently hardcoded in docs/Makefile, can be modified as needed.
• make docs-serve is suitable for development use, supports hot reload
• make docs-build generates complete static website files, output to docs/site directory, can preview final effect locally (open docs/site/index.html)

PDF Manual Build (Pandoc + XeLaTeX)

For generating high-quality developer PDF manuals:

Examples

make pdf
make MONO="JetBrains Mono" pdf      # Override monospace font
make TWOSIDE=1 pdf                   # Two-sided layout (adds -twoside to filename)
make pdf-clean

Dependencies

• pandoc
• XeLaTeX (TexLive)
• Chinese font "Noto Serif CJK SC"
• Monospace font (default DejaVu Sans Mono)
• Optional filter pandoc-crossref

Custom Variables

• MONO Change monospace font
• TWOSIDE Enable two-sided mode when non-empty

Common Issues

• Missing fonts: Install CJK font packages (e.g., fonts-noto-cjk).
• LaTeX errors: Ensure complete XeLaTeX suite is installed (use texlive-full if necessary).
• Missing cross-references: Confirm pandoc-crossref is in PATH.

Output: ucagent-doc.pdf can be distributed with version releases.

Get More Help

• 📚 UCAgent Online Documentation
• 🚀 Quick Start Guide
• 🔧 Custom Configuration
• 🛠️ Tool List
• 💬 GitHub Issues

Contributing

Issues and Pull Requests are welcome!