@openfinclaw/cli
v0.6.0
Published
One-stop quant-trading AI agent for any coding assistant — MCP server + CLI
Readme
@openfinclaw/cli
One-stop quant-trading AI agent for any coding assistant — MCP server + CLI
Market data · analysis · deep reports · strategy generation · backtest · paper trading — all via DeepAgent, across 20+ AI agent platforms.
Try DeepAgent online · Get an API Key · Quick Start · Supported Platforms · GitHub
🧪 Try DeepAgent in your browser first → https://hub.openfinclaw.ai/en/chat — no install required.
What is it?
OpenFinClaw plugs a one-stop quant-trading agent into any MCP-compatible coding assistant. Its core is DeepAgent: a remote multi-agent service that fetches market data, analyses it, writes deep reports, generates strategies, backtests them and proposes paper-trade plans — from a single natural-language prompt. Reach it from Claude Code, Claude Desktop, Cursor, VS Code Copilot, Hermes, and 20+ other platforms. Also runs as a standalone terminal CLI.
Tool groups
| Group | Tools | Tokens | Purpose |
|---|---|---|---|
| deepagent | fin_deepagent_* (14 tools) | ~1,400 | One-stop quant agent — research · analysis · reports · strategy gen · backtest · paper trade |
| strategy | skill_publish · skill_validate · skill_fork · skill_leaderboard · skill_get_info · skill_list_local · skill_publish_verify | ~1,000 | Advanced local FEP v2.0 workflow: publish / fork / validate strategy packages, query Hub leaderboards |
Load only what you need with --tools=deepagent,strategy (either subset).
Quick Start
💡 Not ready to install? Try DeepAgent online and decide after.
1. Interactive setup (recommended)
npx @openfinclaw/cli initThe wizard will:
- Ask for your DeepAgent API Key (required for the
deepagentgroup) - Optionally ask for a Hub API Key (
fch_..., only if you enable thestrategygroup) - Let you pick tool groups
- Auto-detect installed platforms (by app bundles, user data dirs, CLI on
PATH, or existing MCP config) - Write MCP config to each selected platform
- Save keys to
~/.openfinclaw/config.json(chmod 600 on Unix) so the terminal CLI works withoutexport
init also supports scripted, non-interactive use (useful on Windows / CI where raw-mode prompts don't work):
npx @openfinclaw/cli init --yes \
--platforms cursor,claude-code \
--tool-groups deepagent,strategy \
--api-key fch_xxx2. Manual MCP configuration
Drop this into your agent platform's MCP config file (path varies per platform — see below):
{
"mcpServers": {
"openfinclaw": {
"command": "npx",
"args": ["@openfinclaw/cli", "serve"],
"env": {
"OPENFINCLAW_API_KEY": "fch_..."
}
}
}
}单把 fch_ 即可同时驱动 strategy 和 deepagent 两个工具组——deepagent 经 Hub Gateway 鉴权,strategy 直接打 commons-hub。
3. Terminal CLI
For repeated terminal use, install globally so you can call openfinclaw directly instead of prefixing every command with npx:
npm install -g @openfinclaw/cli # or: pnpm add -g @openfinclaw/cli
openfinclaw doctor # confirms the install + keysNot installing globally? Every example below also works prefixed with
npx -y @openfinclaw/cli(slower first run).
# DeepAgent — real token-by-token streaming in the terminal, covers the whole quant flow
openfinclaw deepagent research "Research NVDA last 90 days, propose a momentum strategy, backtest 1y, paper-trade plan"
openfinclaw deepagent skills
openfinclaw deepagent backtests
openfinclaw deepagent packages
openfinclaw deepagent download <package_id>
# Strategy management (Hub key required)
openfinclaw leaderboard --limit 10
openfinclaw strategy-info <uuid>
openfinclaw fork <uuid>
openfinclaw validate ./my-strategy
openfinclaw publish ./my-strategy.zip
# Diagnostics (works with any subset of keys)
openfinclaw doctorOne unified
fch_key. Both DeepAgent and Hub strategy commands shareOPENFINCLAW_API_KEY. DeepAgent traffic is routed through the Hub Gateway which enforces fch_ key auth + multi-tenant isolation.doctorstill renders diagnostics if the key is absent.
Supported Platforms
| Category | Platforms | |---|---| | Chat | Claude Desktop, ChatGPT, Chatbox, LM Studio | | IDEs | Claude Code, VS Code (Copilot), Cursor, Windsurf, JetBrains Junie, Zed, Cline, Continue.dev, Roo Code | | CLI Agents | Codex (OpenAI), OpenCode, Amazon Q CLI | | Frameworks | Hermes Agent, BeeAI, Swarms | | AI Agents | OpenClaw, NanoClaw | | Other | v0 (Vercel), Postman, Amp (Sourcegraph) |
Platform config snippets
{
"mcpServers": {
"openfinclaw": {
"command": "npx",
"args": ["@openfinclaw/cli", "serve"],
"env": {
"OPENFINCLAW_API_KEY": "fch_..."
}
}
}
}{
"mcpServers": {
"openfinclaw": {
"command": "npx",
"args": ["@openfinclaw/cli", "serve"],
"env": {
"OPENFINCLAW_API_KEY": "fch_..."
}
}
}
}{
"servers": {
"openfinclaw": {
"command": "npx",
"args": ["@openfinclaw/cli", "serve"],
"env": {
"OPENFINCLAW_API_KEY": "fch_..."
}
}
}
}mcp_servers:
openfinclaw:
command: "npx"
args: ["@openfinclaw/cli", "serve"]
env:
OPENFINCLAW_API_KEY: "fch_..."{
"mcp": {
"openfinclaw": {
"type": "local",
"command": ["npx", "@openfinclaw/cli", "serve"],
"environment": {
"OPENFINCLAW_API_KEY": "fch_..."
},
"enabled": true
}
}
}Configuration
API Key (one unified fch_...)
| Env var | Used by |
|---|---|
| OPENFINCLAW_API_KEY (fch_..., 68 chars) | Both strategy (commons-hub /api/v1/skill/*) and deepagent (Hub Gateway /api/v1/agent/*) |
doctor renders diagnostics even when the key is absent; every other subcommand requires the key.
Resolution order (highest first):
- CLI flag:
--api-key <key> - Environment variable:
OPENFINCLAW_API_KEY ~/.openfinclaw/config.json(written byopenfinclaw init)
Get a key at hub.openfinclaw.ai, or trial DeepAgent online at https://hub.openfinclaw.ai/en/chat.
Environment variables
| Variable | Description | Default |
|---|---|---|
| OPENFINCLAW_API_KEY | Unified fch_ API key | (required for every authenticated tool) |
| OPENFINCLAW_CONFIG_PATH | Override config file path | ~/.openfinclaw/config.json |
| HUB_API_URL | Hub base URL | https://hub.openfinclaw.ai |
| DEEPAGENT_API_URL | DeepAgent base URL (Hub Gateway) | https://gateway.openfinclaw.ai/api/v1/agent |
| REQUEST_TIMEOUT_MS | HTTP timeout (ms) | 60000 |
| DEEPAGENT_SSE_TIMEOUT_MS | SSE stream timeout (ms) | 900000 (15 min) |
Streaming & long-running tools
DeepAgent research runs can take 3–10 minutes. Two access patterns:
- Terminal CLI (
openfinclaw deepagent research "..."): true token-by-token streaming directly to stdout - MCP clients: the
researchworkflow is split into three calls —fin_deepagent_research_submit(returns ataskIdimmediately),fin_deepagent_research_poll(every 30–60 s for progress),fin_deepagent_research_finalize(fetch full result). The agent drives this loop itself — works consistently across Claude Desktop, Cursor, Hermes, Windsurf, and other MCP clients.
Related
- @openfinclaw/core — Platform-independent library core. Use directly only if you are building a custom MCP wrapper.
- DeepAgent online chat — try it in a browser first.
- GitHub — source, issues, Chinese README.
License
MIT
