The Hyperliquid MCP server that runs real backtests
Keel is a hosted MCP server for Hyperliquid. Add one URL to Claude, ChatGPT, Claude Code or Cursor and your assistant gets 33 tools to search 262 strategy building blocks, build a strategy and backtest it on real Hyperliquid history. There is nothing to install: you sign in once in your browser.
- 1Open Keel in the Claude directory, or search for Keel under Customize → Connectors
- 2Choose Connect
- 3Sign in or sign up on the Keel page that opens (no account needed beforehand)
Sign in in the browser when asked. No API keys. Connect → sign in to Keel → Allow
Or add it by URL
- 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/mcpTwo ways to tell, and either one settles it. A connection that is already live is not improved by signing in again.
In Claude Keel shows as connected once the sign-in has finished (Customize → Connectors, then search for Keel).
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.
- Connect and check your account2
- Find blocks and library strategies6
- Build, copy and restore strategies6
- Run, watch and compare backtests6
- Read strategies, versions and notes9
- Share links and open in the app2
- Help and feedback2
What an MCP server does
MCP, the Model Context Protocol, is an open standard that lets an AI assistant use outside tools. A server publishes a list of tools, each with a name, a short description and typed inputs. Your assistant reads that list when you connect, and calls a tool by name when a request needs it.
That changes what the assistant can tell you. Ask Claude about a Hyperliquid strategy with no server connected and it answers from memory, or writes code you then have to run. With a server connected it can go and get the answer: fetch the data, run the job, and read back what actually happened.
So the useful question about any MCP server is what sits behind its tools. A server that wraps a price endpoint gives your assistant prices. A server with a backtest engine behind it lets your assistant test an idea against years of history and show you the result.
Open-source servers and a hosted engine
Search GitHub for a Hyperliquid MCP server and you will find several open-source projects. Most are small projects, and their READMEs are honest about what they cover. They fall into two groups, and both are good at the job they were built for.
Market data servers
These wrap Hyperliquid's public API: current prices, the order book, open interest and carry rates. They are a quick way to let your assistant look at the market right now, and a good way to learn how an MCP server is put together. They keep no history, so they can't tell you how an idea would have done.
Order servers
These add tools that place and cancel orders, using a key you store in the server's config on your own machine. They suit someone who already knows the rule they want to act on and wants a faster keyboard. There is no backtest and no position sizing behind them, so checking the rule stays your job.
Keel's hosted server
Keel's tools sit on a strategy engine. Your assistant builds a strategy from typed blocks that are checked as they connect, backtests it on Hyperliquid history with fees, slippage and carry, compares versions and links you to the result. It runs on Keel's servers, so there is no code to run and no key on your machine.
| Market data servers | Order servers | Keel MCP | |
|---|---|---|---|
| History to test on | No, live data only | No, live data only | Hyperliquid price bars and hourly carry |
| Builds a strategy | No | No, you bring the rule | Yes, from typed blocks |
| Fees, slippage and carry in results | No results to show | No results to show | Yes, in every backtest |
| Places orders | No | Yes, with a key in its config | No, see the limits below |
| Where it runs | Your machine | Your machine | Hosted, one URL |
| Sign-in | None needed | A key you store | Your browser, once |
Pick by the question you are asking. If you want your assistant to read the market as it is now, a market data server is enough. If you want it to test whether an idea holds up before you act on it, you need history and an engine behind the tools.
One strategy, start to finish
This strategy is in Keel's public library. Your assistant reads it, copies it into your account and changes it with the same tools it uses for a strategy of your own. The canvas is the strategy as the Keel editor shows it, and the chart is its published backtest.
Read it from the top. Price data comes in and splits in two. One branch holds a basket of large coins at equal size. The other watches whether BTC is above its long trend line. A gate lets the basket through only while BTC is above the line, and the last block sizes each position equally. Switch to Code to see the same strategy as text.
Configuration
Factories
Globals(target_timeframe="1d", bar_offset="12h")
Universe(mode="manual", symbols=["BTC", "ETH", "SOL", "HYPE"], market="perp", resolved=["BTC", "ETH", "SOL", "HYPE"], resolved_at="2026-07-06T00:00:00Z")
Execution(
rebalance="buffered",
buffer_threshold=0.1,
buffer_mode="relative",
rebalance_method="to_edge",
)
btc_trend = Pipeline([
AssetSelect(symbols=["BTC"]),
EWMA(window=2),
MarketTrendRegimeFilter(fast_period=2, slow_period=100, index_method="equal_weight"),
], name="btc_trend")
Pipeline([
PriceDataLoader(),
{
"basket": [ConstantForecast(value=10.0)],
"btc_regime": [btc_trend],
},
RegimeGate(signal_key="basket", index_key="btc_regime", condition="index > 0"),
EqualWeightSizer(target_leverage=1.0, max_weight=0.25),
], name="regime_rotation_crypto")
Hyperliquid, Aug 2024 to Sep 2026, after fees, slippage and carry, on the library's default settings. A backtest shows how fixed rules behaved in the past, not what they will do next.
See every setting the library testedFind, build, backtest, review.
Finds the blocks
It searches Keel's 262 building blocks (signals, regime filters, coin selection, position sizing) and reads each one's inputs and settings before it builds anything.
Builds the strategy
It writes the strategy in Keel's strategy language. Every connection is checked as it is saved, so a block that doesn't fit is caught before any backtest runs.
Backtests it on real history
Hyperliquid price bars and hourly carry, with fees and slippage charged. Return, Sharpe and drawdown come back in the conversation, and it can copy the strategy, change one setting and compare the two.
Hands you the result
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.
A step-by-step guide for each assistant
Every client takes the same URL. What differs is where you paste it. Each guide shows the exact menu path or command, what you see when you sign in, and a first request to try.
- ClaudeAdd it from its listing in the app's directoryOpen the guide
- ChatGPTAdd it from its listing in the app's directoryOpen the guide
- Claude CodeOne command in your terminalOpen the guide
- CursorOne entry in its config fileOpen the guide
- CodexOne command in your terminalOpen the guide
- WindsurfOne entry in its config fileOpen the guide
- Any other MCP clientPaste the URL in its settingsOpen the guide
What the server doesn't do
It doesn't place orders
None of the tools sends an order to Hyperliquid. Your assistant builds, tests and reads. Running a strategy on your account happens in the Keel app, where you see exactly what will run before it starts.
It covers Hyperliquid's crypto perps
BTC, ETH, SOL, HYPE and the rest of Hyperliquid's crypto perpetuals. Stock and index perps are HIP-3 markets, and Keel doesn't support them yet. Support is coming soon.
A backtest is still a simulation
Fills are modelled, with Hyperliquid's fee schedule and a slippage estimate for each coin. It tells you how the rules behaved, not what they will do next.
It needs a Keel account
The first time your assistant connects, a Keel page opens to sign in or sign up. Backtests your assistant runs come out of your Keel plan, and it can tell you how many you have left.
Questions about the Keel MCP server
What is the Keel MCP server?
A hosted MCP server that connects any MCP-capable assistant to Keel's strategy engine for Hyperliquid. It runs at https://mcp.usekeel.io/mcp. Add that URL to your client, sign in in the browser when asked, and the 33 tools arrive with the knowledge your assistant needs to use them.
What is an MCP server, in plain words?
A small service that hands an AI assistant a set of tools. The assistant sees each tool's name, description and inputs, and calls one when your request needs it. MCP is the open standard for that handshake, so one server works in Claude, ChatGPT, Cursor and the rest.
How is Keel different from the open-source Hyperliquid MCP servers?
Most open-source servers wrap Hyperliquid's live API: prices, the order book, and in some cases order placement with a key you store. None of them has history or a backtest engine behind it. Keel's server is hosted, builds strategies from typed blocks and backtests them on Hyperliquid history with fees, slippage and carry.
Do I need to install anything?
No. In Claude and ChatGPT you add Keel from its listing in the app's own directory, and in Claude Code, Cursor, Codex or Windsurf one URL is the whole setup. The keel-trade package is the other way in, a local MCP server and a command line tool for file-based work, and it is listed under Other ways to connect on the agents page.
Which clients work with it?
Claude on the web, desktop and phone, ChatGPT, Claude Code, Cursor, Codex and Windsurf, plus any client that accepts a remote MCP URL. The agents page has the steps for each, with what to expect at every step.
What does a backtest run on?
Hyperliquid price bars from 5 minutes to daily, and hourly carry rates, the same series the Keel app uses. Order-flow series are there too (taker flow, large prints, exact VWAP and dollar volume), built from every fill on the venue since March 2025. BTC, ETH and SOL price history starts in June 2024, HYPE in December 2024. Carry is counted hourly and shown apart from the price return.
Can my assistant place orders on Hyperliquid through Keel?
No. The hosted server has no tool that sends orders. It can read how the strategies on your Keel account are doing, and for anything else it gives you a link to the Keel app.
Give your assistant a backtest engine
Paste one URL and ask it to test an idea. You see the strategy and the result before anything else happens.