内容
<div align="center">
# 🧠 ForgeMind
**MCP-native multi-agent AI platform — tools, approvals, traces, evals**
[](LICENSE)
[](https://python.org)
[](docs/mcp.md)
[](src/api)
[](Dockerfile)
[](tests)
**Give it a goal. Planner thinks, Executor acts through MCP tools, Critic verifies — with human approvals and full traces.**
</div>
## Why ForgeMind
2026's AI stack is **agents + MCP + evals**. ForgeMind is a complete, runnable reference implementation of that stack:
- 🔌 **MCP-native** — every tool is exposed as MCP (`tools/list`, `tools/call` over HTTP + stdio); consume any external MCP server too
- 🤝 **Human-in-the-loop** — `auto`, `confirm_tools`, `confirm_all` approval modes for risky tools
- 🔍 **Execution traces** — every run saved as replayable JSON; inspect via CLI, API, or dashboard
- 📊 **Eval harness** — built-in suite with pass-rate reports; add your own cases in one file
- 🖥️ **Live dashboard** — run goals, watch WebSocket events, browse tools and traces
- 🧠 **Any model** — Ollama (free/local), llama.cpp GGUF, or any OpenAI-compatible API
## Architecture
```
Goal → Coordinator (ReAct loop)
├── PlannerAgent → thought + tool_calls (JSON)
├── ApprovalGate → auto / confirm_tools / confirm_all
├── ExecutorAgent → ToolRegistry.execute()
│ ├── builtin: calculator, time, echo, word_count
│ └── risky: write_file, run_shell, http_request
├── CriticAgent → accept / reject final answer
└── Tracer → .forgemind/traces/<run_id>.json
Interfaces: CLI · FastAPI (/v1/runs, /v1/tools, /v1/approvals, /v1/traces)
MCP endpoint (/v1/mcp) · WebSocket (/v1/stream) · Dashboard (/dashboard)
```
## Quick Start
```bash
pip install -e ".[dev]"
# Needs a model backend — Ollama is free and local:
ollama pull qwen2.5-coder:7b
# Run a goal
forgemind run "What is (12 + 8) * 3? Reply with just the number."
# Start the API + dashboard
forgemind serve
# → http://127.0.0.1:8080/dashboard
# Run evals
forgemind evaluate
```
## API Examples
```bash
# Start a run
curl -X POST localhost:8080/v1/runs \
-H "Content-Type: application/json" \
-d '{"goal": "Echo back exactly: hello-forge", "max_steps": 8}'
# List tools (MCP)
curl localhost:8080/v1/tools
# Call a tool via MCP
curl -X POST localhost:8080/v1/mcp \
-H "Content-Type: application/json" \
-d '{"jsonrpc":"2.0","id":1,"method":"tools/call",
"params":{"name":"calculator","arguments":{"expression":"3*7"}}}'
# Pending approvals for a run
curl localhost:8080/v1/approvals/<run_id>
```
## Docker
```bash
docker compose up --build
# API on :8080, data in the forgemind-data volume
```
## CLI Reference
| Command | Purpose |
|---|---|
| `forgemind run "<goal>"` | Execute one goal end to end |
| `forgemind serve` | Start API + dashboard |
| `forgemind evaluate` | Run eval suite, print Markdown report |
| `forgemind tools` | List registered MCP tools |
| `forgemind trace <run_id>` | Print a stored trace |
Env vars: `FORGEMIND_BACKEND`, `FORGEMIND_MODEL`, `FORGEMIND_API_KEY`,
`FORGEMIND_APPROVALS`, `FORGEMIND_MAX_STEPS`, `FORGEMIND_PORT`.
## License
MIT — see [LICENSE](LICENSE)
<div align="center">
<b>Built by <a href="https://github.com/kamalesh404">Kamalesh</a></b>
</div>
MCP 市场