MCP 服务器

OpenFinClaw CLI

Enables quant research, strategy generation, backtesting, and paper trading from natural language prompts, integrating with AI agents via an MCP server.

README

English | 中文

Your quant research team, in one prompt.

Research · strategy · backtest · paper trade — ship a complete quant workflow from a single natural-language prompt, inside Claude Code, Cursor, and 20+ AI agents.

🚀 Try it in 60 seconds — zero install

Run a full research → strategy → backtest loop in your browser. No install, no API key, real market data.

Quick Start · Example Prompts · Community · Platforms · vs. other tools

</div>

What you get


🧠 DeepAgent analysis skills	60+ built-in — technical · fundamental · sentiment · risk · timing · factor
🌍 Markets covered	5 — US equities · A-shares · HK · Crypto · Forex
🤖 Works with	20+ AI platforms — Claude Code · Cursor · VS Code · Hermes · Windsurf · Codex …
🔄 End-to-end flow	research → strategy → backtest → paper trade → publish to the leaderboard
⚡ How you interact	streaming token-by-token in the terminal · MCP tool calls · in-browser playground

<img src="imgs/deepagent-backtest-metrics.png" alt="DeepAgent backtest result — Tesla Bollinger Bands" width="620"> Live output from <code>openfinclaw deepagent research</code> — one prompt: research → strategy → backtest → metrics.

Example Prompts

Copy-paste any of these into openfinclaw deepagent research "…" (or drop them straight into your AI agent). Each one runs the full research → strategy → backtest loop.

📈 Technical analysis

Find RSI divergence signals on NVDA in the last 6 months, then backtest them.
Compare a Bollinger Bands strategy on TSLA vs AAPL over 1 year — which wins?
Screen the S&P 500 for golden-cross signals this month.

📊 Fundamentals & macro

Pull Apple's last 8 quarters of revenue, margins, and guidance. Summarize the trend.
What's driving the NVDA move this quarter — earnings, guidance, or narrative?
Compare AMD / INTC / NVDA on growth, margin, and valuation.

🎯 Strategy generation

Design a momentum strategy on US mega-cap tech. Backtest 2y. Tell me where it breaks.
Write a mean-reversion strategy on BTC and show drawdown behavior through 2022.
A-shares 沪深 300 日内轮动策略，年化目标 15%，最大回撤 < 10%。

🧪 Backtest & stress-test

Backtest a 50/200 SMA crossover on SPY from 2015. Include costs and slippage.
Stress-test my forked strategy against the 2020 and 2022 crashes.

Want a ready-made one? Run openfinclaw leaderboard to browse the community's highest-ranked strategies, then fork any of them.

Quick Start

💡 Want to see it in action before installing? Try DeepAgent in your browser first.

60-second onboarding

npx @openfinclaw/cli@latest install               # wizard + MCP configs + Skill registration + doctor
openfinclaw deepagent +research "盘点 BTC 周线"     # streaming research / strategy / backtest

install runs the interactive wizard, writes MCP configs to every detected AI agent, persists your fch_ key to ~/.openfinclaw/config.json (chmod 600 on Unix), drops a SKILL.md under ~/.claude/skills/openfinclaw/ so Claude Code / Cursor auto-trigger on keywords like quant / backtest / 量化, and finishes with a connectivity check.

Non-interactive / CI:

npx @openfinclaw/cli@latest install --yes \
  --platforms cursor,claude-code --tool-groups deepagent,strategy \
  --api-key fch_xxx --register-skill

Just the wizard, no SKILL.md registration, no doctor: npx @openfinclaw/cli init.

CLI quick reference

A single fch_ key drives both DeepAgent and the strategy group. Resolution order: --api-key → OPENFINCLAW_API_KEY → ~/.openfinclaw/config.json.

Group	Commands
DeepAgent	`deepagent +research "<query>"`, `deepagent health`, `deepagent skills`, `deepagent threads`, `deepagent messages`, `deepagent backtests`, `deepagent packages`, `deepagent download`
Strategy	`leaderboard`, `strategy-info`, `fork`, `list-strategies`, `validate`, `publish`, `publish-verify`
Raw	`api GET <path>` · `api POST <path> --json '<body>'` — direct Hub Gateway call, auth pre-attached
System	`install` · `init` · `skill-install` · `serve` · `doctor` · `update` · `examples`

+verb (e.g. deepagent +research) is the human-friendly streaming path; the bare verbs and MCP-only atomic triplet research_submit / research_poll / research_finalize are for agents/scripts. Run openfinclaw --help for the full surface.

Sample DeepAgent output — one prompt → strategy definition + backtest metrics + per-trade P&L + improvement notes:

Community: leaderboard → fork → publish

OpenFinClaw ships with a community strategy exchange. Browse what others are running, copy any strategy locally, tweak it, and publish back — think of it as a Hugging Face for quant strategies.

openfinclaw leaderboard --limit 20          # Browse top-ranked strategies
openfinclaw strategy-info <id>              # See how a strategy performs
openfinclaw fork <id>                       # Copy to ./strategies/<slug>
# ... edit strategy.py, tweak fep.yaml ...
openfinclaw validate ./strategies/<slug>    # Pre-flight FEP v2.0 check
openfinclaw publish ./my-strategy.zip       # Ship to the leaderboard
openfinclaw publish-verify --submission-id <id>   # Track backtest progress

Every published strategy is backtested server-side and ranked by live-market-equivalent returns — no self-reported numbers.

Supported Platforms

OpenFinClaw works with any MCP-compatible agent platform:

Category	Platforms
Chat	Claude Desktop, Claude.ai, ChatGPT, Chatbox, LM Studio
IDEs	Claude Code, VS Code (Copilot), Cursor, Windsurf, JetBrains Junie, Zed, Cline, Continue.dev
CLI Agents	Codex (OpenAI), OpenCode, Amazon Q CLI
Frameworks	Hermes Agent, BeeAI, Swarms
AI Agents	OpenClaw, NanoClaw
Other	v0 (Vercel), Postman, Roo Code, Amp (Sourcegraph)

Platform Config Examples

<details> <summary>Claude Code — <code>~/.claude/settings.json</code></summary>

{
  "mcpServers": {
    "openfinclaw": {
      "command": "npx",
      "args": ["@openfinclaw/cli", "serve", "--tools=deepagent,strategy"],
      "env": {
        "OPENFINCLAW_API_KEY": "fch_xxx"
      }
    }
  }
}

</details>

<details> <summary>Cursor — <code>.cursor/mcp.json</code></summary>

{
  "mcpServers": {
    "openfinclaw": {
      "command": "npx",
      "args": ["@openfinclaw/cli", "serve", "--tools=deepagent,strategy"],
      "env": {
        "OPENFINCLAW_API_KEY": "fch_xxx"
      }
    }
  }
}

</details>

For other platforms (VS Code, Hermes, Windsurf, Zed, OpenClaw, Junie, Trae, …), see configs/ for ready-to-copy templates. The shape is the same — only the host key (servers vs mcpServers vs context_servers) and config path differ.

Tool Groups & Context Optimization

Load only what you need to save tokens: serve --tools=deepagent (~1,400 tk) or serve --tools=strategy (~1,000 tk), or omit --tools for both.

Group	Tools
`deepagent`	14 remote-agent tools — `fin_deepagent_health` / `_skills` / `_research_submit` / `_research_poll` / `_research_finalize` / `_status` / `_cancel` / `_threads` / `_messages` / `_backtests` / `_backtest_result` / `_packages` / `_package_meta` / `_download_package`
`strategy`	7 local FEP v2.0 tools — `strategy_publish` / `strategy_validate` / `strategy_fork` / `strategy_leaderboard` / `strategy_get_info` / `strategy_list_local` / `strategy_publish_verify`

Environment Variables

Only one is required:

Variable	Description
`OPENFINCLAW_API_KEY`	Unified `fch_` key. Drives both strategy (Hub) and deepagent (Hub Gateway). Falls back to `~/.openfinclaw/config.json` if unset. Get a key at hub.openfinclaw.ai.

Advanced overrides (rarely needed): OPENFINCLAW_CONFIG_PATH, HUB_API_URL, DEEPAGENT_API_URL, REQUEST_TIMEOUT_MS, DEEPAGENT_SSE_TIMEOUT_MS — see packages/core/src/config.ts.

Development

git clone https://github.com/mirror29/openfinclaw-cli.git && cd openfinclaw-cli && pnpm install && pnpm build
OPENFINCLAW_API_KEY=<fch_...> node packages/cli/dist/index.js doctor   # smoke test

Monorepo: @openfinclaw/core (zero-dep business logic) + @openfinclaw/cli (MCP server + terminal CLI + install wizard).

License

MIT