lighter-mcp
A portable MCP server that exposes Lighter trading to any MCP-capable agent with safety-first features like mode-based gating, two-step confirmations, and audit logging.
README
Lighter MCP Toolkit
Read this first: live orders and withdrawals on Lighter are irreversible. Read DISCLAIMER.md before flipping
mode = "live"ormode = "funds".
A portable Model Context Protocol server that exposes Lighter trading to any MCP-capable agent — Cursor, Claude Code, Claude Desktop, Codex, OpenClaw, and generic MCP clients — with safety as a first-class concern, not an afterthought.
The server wraps the official
lighter-agent-kit
Python scripts via subprocess and adds:
- Typed Pydantic input schemas with regex-validated symbols and asset codes (no command-line argument injection).
- Mode-based safety gates (
readonly→paper→live→funds): the tool catalog grows monotonically; live and funds tools physically do not exist in the catalog until you opt in. - Two-step confirmations for every fund-loss path: first call returns a preview + single-use token, second call with the token executes. Tokens are bound to tool name + argument digest and TTL-limited.
- Atomic, UTC-keyed daily-notional ledger that fail-safes to
cap exhaustedif its file is corrupt — never silently resets to 0. - Append-only audit log with sanitized argv/result, intra-process
lock + POSIX
flock, and soft-fail on disk errors. - Streamable-HTTP transport that refuses to bind non-loopback hosts
unless
--allow-remoteis passed (the server has no built-in auth).
Why not just use the kit directly?
The kit is a CLI for a human. This server is the safety harness for a non-human caller. Every call is schema-validated, gate-checked, audited, and (for write paths) preview-confirmed before any subprocess runs against the exchange. The agent literally cannot reach the kit without going through these layers.
Quick start
One command, end to end:
pipx install lighter-mcp && lighter-mcp init
lighter-mcp init clones the upstream lighter-agent-kit into
~/.lighter/lighter-agent-kit, writes a default readonly config to
~/.lighter/lighter-mcp/config.toml, and auto-wires every MCP-capable
agent it finds locally — Cursor, Claude Code, Claude Desktop, Codex —
including slash-commands, the lighter-trader sub-agent, and the
post-trade hook. Restart your agent and try:
/lighter-status
That's it. Live trading stays OFF until you flip it on in the config and
re-run lighter-mcp doctor.
Other ways to install
<details> <summary><b>uvx (zero install, ephemeral)</b></summary>
uvx lighter-mcp init
uvx runs the published wheel from a cached venv. Good for one-off
exploration; for steady-state use prefer pipx install so the agent
config can point at a stable executable path.
</details>
<details> <summary><b>Docker (GHCR)</b></summary>
docker pull ghcr.io/iamgoatedaf/lighter-mcp:0.1.0
docker run --rm -it -v ~/.lighter:/data \
ghcr.io/iamgoatedaf/lighter-mcp:0.1.0 \
lighter-mcp init --no-scaffolds
The image ships with the kit pre-cloned at /opt/lighter-agent-kit and
runs lighter-mcp serve on :8791 by default.
</details>
<details> <summary><b>From source (development)</b></summary>
git clone https://github.com/iamgoatedaf/lighter-mcp
cd lighter-mcp
python3 -m venv .venv && .venv/bin/pip install -e ".[dev,http]"
.venv/bin/lighter-mcp init
</details>
After any path above, lighter-mcp doctor should print a green health envelope.
Modes and tools
| Mode | Reads | Paper trading | Live trading | Transfers / Withdrawals | Tools |
|---|---|---|---|---|---|
| readonly | ✅ | — | — | — | 18 |
| paper | ✅ | ✅ | — | — | 29 |
| live | ✅ | ✅ | gated | — | 37 |
| funds | ✅ | ✅ | gated | gated | 39 |
live and funds additionally require explicit [live] / [funds]
config blocks with allowlists, notional caps, leverage caps, and
confirmation flags. See docs/configuration.md.
Slash commands and sub-agent (Cursor / Claude Code / Codex)
The adapters ship a pre-built UI layer driven by the same MCP tools:
| Command | What it does |
|---|---|
/lighter-status |
Active mode, daily-notional remaining, equity, open orders. |
/lighter-positions |
Positions with entry / mark / PnL / liq distance. |
/lighter-kill |
Panic close: cancel_all + close_all, two-step confirm with the literal word confirm. |
/lighter-paper |
Switch active mode to paper. Prompts before editing config. |
/lighter-audit |
Tail the audit log with filters (last N hours, only failures, by tool). |
Plus a dedicated lighter-trader sub-agent with a narrow
lighter_*-only tool budget that cannot edit source files and refuses to
chain a kill without an explicit confirmation word.
The single source of truth for these prompts lives in
adapters/_shared/. Per-platform folders are
symlinks back into it.
Agent adapters
| Platform | MCP tools | SKILL / system prompt | Slash commands | Sub-agent | Hook | Auto-install |
|---|---|---|---|---|---|---|
| Cursor | ✅ | rule (.mdc) |
✅ | ✅ | ✅ | lighter-mcp init --agents cursor |
| Claude Code | ✅ | SKILL.md | ✅ | ✅ | ✅ | lighter-mcp init --agents claude-code |
| Claude Desktop | ✅ | SKILL paste | ❌ | ❌ | ❌ | lighter-mcp init --agents claude-desktop |
| Codex | ✅ | SKILL.md | ✅ | ✅ | ✅ | lighter-mcp init --agents codex |
| OpenClaw / Telegram | ✅ | mapping doc | bot-side | ❌ | bot-side | manual — see adapters/openclaw/ |
| Generic MCP | ✅ | by hand | — | — | — | see adapters/generic/ |
lighter-mcp init (no --agents) detects all installed agents at once.
Per-platform READMEs explain exactly which UI surfaces exist for that
agent and which require workarounds.
Examples
End-to-end walkthroughs in examples/:
paper-demo.md— first paper trade, step-by-step.guarded-live-order.md— full two-step confirm flow onlighter_live_market_order.funding-scan.md— read-only scan of funding rates across symbols.live-pnl-watcher/— two-pane terminal workflow: a WebSocket-driven PnL view that hot-reloads onpaper-state.jsonchanges, plus atrade.shCLI wrapper for long/short/status while the watcher is running.
Hosted documentation site — website/ is a Mintlify
project. Push to GitHub and Mintlify auto-deploys the site at
https://lighter-mcp.mintlify.app (or your custom domain). Local
preview: cd website && npm run dev.
Markdown deep-dives in docs/ (legacy / source material for
the site):
getting-started.mdconfiguration.md— every key in the TOML.audit-log.md— record format, redaction rules, retention strategy.security.md— companion to top-levelSECURITY.md.
Development
python3.11 -m venv .venv
.venv/bin/pip install -e ".[dev,http]"
.venv/bin/ruff check .
.venv/bin/pytest -q # 53 tests, fully hermetic by default
Tests that need a real lighter-agent-kit checkout are auto-skipped
unless LIGHTER_KIT_PATH is set. See CONTRIBUTING.md
for the contribution workflow and the bar for changes that touch
safety.py, confirmations.py, or any lighter_live_* /
lighter_funds_* tool.
Compatibility notes
- Tested against
lighter-agent-kitas of April 2026. The wrapper relies on the kit'squery.py/trade.py/paper.pyscript entry points and their JSON stdout contracts — pin to a known-good kit commit if you fork. - Python ≥ 3.10. Built and tested on macOS and Ubuntu; Windows works for
the server itself, but the per-platform adapters use symlinks — run
adapters/_shared/sync.shto materialize them as real files.
Project status
Alpha (0.1.0). Tool surface is stable enough to use, but expect
schema-level breaking changes between minor versions until 1.0. See
CHANGELOG.md.
License
MIT — see LICENSE.
推荐服务器
Baidu Map
百度地图核心API现已全面兼容MCP协议,是国内首家兼容MCP协议的地图服务商。
Playwright MCP Server
一个模型上下文协议服务器,它使大型语言模型能够通过结构化的可访问性快照与网页进行交互,而无需视觉模型或屏幕截图。
Magic Component Platform (MCP)
一个由人工智能驱动的工具,可以从自然语言描述生成现代化的用户界面组件,并与流行的集成开发环境(IDE)集成,从而简化用户界面开发流程。
Audiense Insights MCP Server
通过模型上下文协议启用与 Audiense Insights 账户的交互,从而促进营销洞察和受众数据的提取和分析,包括人口统计信息、行为和影响者互动。
VeyraX
一个单一的 MCP 工具,连接你所有喜爱的工具:Gmail、日历以及其他 40 多个工具。
graphlit-mcp-server
模型上下文协议 (MCP) 服务器实现了 MCP 客户端与 Graphlit 服务之间的集成。 除了网络爬取之外,还可以将任何内容(从 Slack 到 Gmail 再到播客订阅源)导入到 Graphlit 项目中,然后从 MCP 客户端检索相关内容。
Kagi MCP Server
一个 MCP 服务器,集成了 Kagi 搜索功能和 Claude AI,使 Claude 能够在回答需要最新信息的问题时执行实时网络搜索。
e2b-mcp-server
使用 MCP 通过 e2b 运行代码。
Neon MCP Server
用于与 Neon 管理 API 和数据库交互的 MCP 服务器
Exa MCP Server
模型上下文协议(MCP)服务器允许像 Claude 这样的 AI 助手使用 Exa AI 搜索 API 进行网络搜索。这种设置允许 AI 模型以安全和受控的方式获取实时的网络信息。