MCP for Hyperliquid, with a real backtest engine behind it
Model Context Protocol lets an assistant call tools instead of scraping a screen. The Keel MCP server gives Claude, ChatGPT, Claude Code, Cursor and any MCP client 29 tools on Hyperliquid: a typed strategy DSL, a backtest engine on real Hyperliquid history, and the result read back to you. One URL, nothing to install.
- 1Customize → Connectors → Add custom connector
- 2Paste https://mcp.usekeel.io/mcp and choose Connect
- 3Sign in or sign up on the Keel page that opens (no account needed beforehand)
https://mcp.usekeel.io/mcpSign in in the browser when asked. No API keys. Connect → sign in to Keel → Allow
Two ways to tell, and either one settles it — a connection that is already live is not improved by signing in again.
In Claude Keel is listed where you added it (Customize → Connectors → Add custom connector) once the sign-in has finished.
Ask the agent keel_account_status answers with the Keel account the session is signed in as, which nothing but a live connection can return. If the agent says it has no Keel tools, the sign-in did not finish.
Use keel_account_status and tell me which Keel account this session is signed in as.Not yet verified against a Claude version. These steps follow Anthropic's own documentation, read on 2026-09-16.
Most Hyperliquid MCP servers stop at a price feed.
Price and account wrappers
The open-source Hyperliquid MCP servers wrap the exchange API: a price endpoint, an open-interest endpoint, maybe order placement. The agent can read a number; it cannot test an idea.
Scripts the assistant writes
A one-off Python script against the exchange API has no portfolio simulation, no fee schedule, no carry accounting and no version history. Every idea starts from a blank file.
The Keel MCP server
Your agent composes against 218 typed components, backtests on Hyperliquid 15-minute bars and hourly carry with real fees, and forks and compares versions — the same engine the Keel app uses.
Then the app
The agent builds, tests and reads. When you want to run a strategy on your account, it hands it to the Keel app and you take it from there.
Questions about MCP on Hyperliquid
What is an MCP server?
Model Context Protocol is the open standard that lets assistants like Claude, ChatGPT and Cursor call outside tools through typed schemas instead of scraping a screen. An MCP server publishes those tools; the assistant calls them by name inside your conversation.
How do I connect Claude Code to the Keel MCP server?
Pick Claude Code in the box above and run its one command, then type /mcp inside Claude Code and choose Authenticate. Your browser opens the Keel sign-in, and the 29 tools appear in the session. Cursor, Codex and Windsurf have their own tab with the same URL.
What data does a backtest run on?
Hyperliquid price bars from 5 minutes up and hourly carry rates, the same series the Keel app uses, plus order-flow series (taker flow and CVD, whale prints, exact VWAP, TWAP share, dollar volume) built from 1-minute bars derived from every fill on the venue since March 2025. BTC, ETH and SOL price history starts in June 2024 with carry from late July 2024; HYPE starts 2024-12-05. Carry accrues hourly into the equity curve and is shown separately from the price return.
How is Keel different from the Hyperliquid MCP servers on GitHub?
Those expose a handful of read-only price and account endpoints. There is no backtest engine behind them and no typed component graph, so the agent stops at a number. Keel is the engine those servers would call if it existed: compose, backtest, fork, compare, share.