ChatGPT and the OpenAI stack — three routes to the same engine.
ChatGPT Connectors and the OpenAI Codex CLI both speak MCP. Custom GPTs call the Keel REST API through Actions. All three routes drive the same deterministic Hyperliquid backtest engine — composition, backtest, and non-custodial live execution. Pick the route that matches your workflow.
ChatGPT Connectors, the Codex CLI, and custom GPTs each connect differently.
The OpenAI ecosystem reaches Keel three ways. Two of them use MCP directly; one uses REST through Actions. None of them put the LLM in the trade-execution loop — the compiled strategy runs on its own schedule.
In-chat MCP access. Register the hosted Keel MCP endpoint as a Connector (developer mode) and ChatGPT can call its tools — compose a strategy, run a backtest, pull a tearsheet — without leaving the chat window and with nothing installed locally.
Best for users who live in the ChatGPT app and want strategy authoring as a conversation. The agent calls Keel tools by name; results stream back inline.
A terminal-native developer agent with native MCP support. Add Keel to your Codex config, run codex in any project, and the agent can search components, compose strategies, run backtests, and fork variants — same protocol Claude Code and Cursor use over stdio.
Best for users whose workflow lives in the terminal. Pair with your editor of choice. The CLI's agent loop handles long-running backtests and multi-step composition cleanly.
Lightweight REST-based integration. Build a custom GPT, import the Keel OpenAPI spec, add an API key. The GPT can pull funding rates, summarize backtests, and check account state. Doesn't require installing a CLI or running anything locally.
Best for read-heavy flows and shareable assistants. Anyone with the GPT link can use it. Less suited to long-running strategy authoring than the MCP routes, but powerful for analyst and review work.
Pick a route. Run the snippet. The agent has access.
Route A registers the hosted Keel MCP endpoint as a ChatGPT Connector. Route B installs the local stdio server for the Codex CLI. Route C imports the Keel OpenAPI spec into a custom GPT for REST-based flows.
ChatGPT Connectors (remote MCP)
In-chat MCP with nothing installed. Register the hosted Keel endpoint as a Connector and the agent has direct tool access.
OpenAI Codex CLI
Terminal-native MCP. Register the local stdio server with Codex, then run Codex in your project.
Custom GPT with Actions
REST-based. Build a custom GPT, import the Keel OpenAPI spec, add an API key. Good for read-only and shareable flows.
Common questions
How do I connect ChatGPT to Hyperliquid?
Three routes today. (1) ChatGPT Connectors — register the hosted Keel MCP endpoint (mcp.usekeel.io/mcp) as a Connector in developer mode and call its tools from chat, nothing installed. (2) The OpenAI Codex CLI — a developer agent that speaks MCP natively over stdio via pipx install keel-trade. (3) A custom GPT with Actions — imports the Keel OpenAPI spec and calls REST endpoints. Pick the route that fits your workflow; each one drives the same engine.
Does ChatGPT support MCP?
Yes, on two surfaces. ChatGPT Connectors take a remote MCP server URL — the hosted Keel endpoint plugs straight in. The OpenAI Codex CLI supports MCP stdio servers the same way Claude Code, Cursor, and Windsurf do. Custom GPTs use Actions instead — REST calls against an OpenAPI spec — which is a different protocol but still gets you to the same Keel surface.
Which route should I pick?
ChatGPT Connectors if you live in the ChatGPT app and want in-chat strategy authoring with nothing installed. The Codex CLI if your work is terminal-driven and you want a developer agent next to your editor, with local file-based workspaces. A custom GPT with Actions if you want a lightweight read-only assistant — funding lookups, tearsheet summaries, account checks.
Can ChatGPT or Codex trade my Hyperliquid account autonomously?
No. With any of the three routes, the LLM helps you compose a strategy and run a backtest. The compiled strategy artifact is what trades — it runs on its own schedule, on the same execution path the backtest used. The LLM is not in the runtime loop. Live execution requires explicit OAuth scope and a separate local arming step; non-custodial throughout via Hyperliquid native delegated signing.
Can I trigger a Keel backtest from ChatGPT?
Yes from all three routes. ChatGPT Connectors and the Codex CLI both call the backtest tool directly over MCP and stream results back into the conversation. A custom GPT with Actions calls the same backtest endpoint over REST — slightly more polling involved for long-running jobs, but you get the same artifact and the same share URL.
Where do I get the OpenAPI spec or the MCP connection details?
For ChatGPT Connectors, use the hosted endpoint URL mcp.usekeel.io/mcp — no install. For the Codex CLI, install the Keel package (pipx install keel-trade) and register keel mcp serve — the per-host snippet matrix is documented at /docs/sdk/agent-setup and the full runbook at /agents. For a custom GPT, import the Keel OpenAPI spec and add an API key from your account settings.
Keep exploring
The Keel MCP server
Product page for the MCP. Install path, tool surface, browser OAuth, and the full per-host config matrix.
Claude × Hyperliquid
Sibling page for Claude Code, Claude Desktop, and Cursor. Same engine, different host.
Build a trading strategy with an agent
Step-by-step tutorial covering composition, backtest, and non-custodial live deploy through any MCP-capable agent.