AI工作流平台

Requirements

• A supported Python 3 environment
• A virtual environment is strongly recommended
• Credentials for at least one model provider, unless you configure a local provider
• On some Linux systems, libgl-dev may be required

What you can build

MindRoot is intended for real applications rather than one narrow agent pattern. Examples include:

• Internal research and operations assistants
• Document extraction and report-generation systems
• Knowledge-base and RAG applications
• Browser, desktop, and shell automation
• Background and bulk-processing workflows
• Voice and multimodal agents
• Database-backed business assistants
• Rich data viewers, dashboards, and generated workspaces
• Domain-specific products with a completely custom UI

Existing plugins cover integrations such as Anthropic, OpenAI, OpenRouter, Gemini, Groq, DeepSeek, Cerebras, Fireworks, Together AI, Deepgram, image and video generation, browser and computer control, SQL databases, Supabase, file and Office-document operations, MCP, persistent memory, knowledge bases, job queues, and custom UI components.

The plugin ecosystem changes faster than this README. Use the admin plugin index or install a compatible plugin directly from GitHub to inspect the currently available integrations.

1. Install

python -m venv .venv
source .venv/bin/activate
pip install mindroot

or OPENAI_API_KEY, or credentials for another installed provider

Optional email verification:

See the SMTP plugin documentation for mail configuration.

Quick start

Gallery

2. Configure

Set a secret and the credentials required by your chosen provider:

```bash export JWT_SECRET_KEY="replace-with-a-long-random-value" export ANTHROPIC_API_KEY="..."

4. Configure an agent

Open /admin and:

1. Install a model-provider plugin.
2. Configure its required environment variables.
3. Create or select an agent.
4. Enable only the commands that agent should be allowed to use.
5. Restart MindRoot after installing or changing plugins when prompted.

You now have an agent accessible through the web interface and programmatically.

REST API

Run an agent as a long-running task:

curl -X POST \
  "http://localhost:8010/task/Assistant?api_key=${MINDROOT_API_KEY}" \
  -H "Content-Type: application/json" \
  -d '{"instructions":"Inspect this request, use the enabled tools, and return a concise report."}' \
  --max-time 300

A successful response includes the final result, task trace, and conversation log identifier:

{
  "status": "ok",
  "results": "Final textual or structured result",
  "full_results": [
    {
      "cmd": "some_command",
      "args": {},
      "result": "..."
    }
  ],
  "log_id": "..."
}

See API documentation for authentication, task-agent configuration, endpoint behavior, and additional examples.

Python SDK

pip install mrsdk

from mrsdk import MindRootClient

client = MindRootClient(
    api_key="your_api_key",
    base_url="http://localhost:8010",
)

result = client.run_task(
    agent_name="Assistant",
    instructions="What is the square root of 256? Show your work.",
)

print(result["results"])

See the mrsdk repository for SDK details and task-trace access.

Admin interface

The plugin model

A MindRoot plugin is a normal installable Python package that can contribute one or more layers of an application:

my_plugin/
├── plugin_info.json          # Metadata, commands and services
├── pyproject.toml
└── src/my_plugin/
    ├── mod.py                # Agent commands and internal services
    ├── router.py             # Optional FastAPI routes
    ├── static/               # JavaScript, CSS and other assets
    ├── templates/            # Plugin-owned pages
    ├── inject/               # Add content to existing template blocks
    └── override/             # Replace existing template blocks

Full-stack plugins

Plugins are not limited to tools. They can add FastAPI routes and complete frontend experiences using Jinja2 and Lit Web Components. The standard chat UI exposes command lifecycle events such as partial output, running state, final results, media, and completion. A plugin can register a renderer for its command and turn structured output into a chart, table, editor, approval form, or other interactive interface.

This allows domain applications to live with the agent rather than maintaining a disconnected frontend and orchestration stack.

See Plugin documentation for package structure, decorators, routes, template injection, component integration, SSE events, pipelines, and development guidance.

Plugin management

HeyGen integration