Skip to content
Cirrus home
Sign in

Guide

Connect an AI agent

Connect Claude, ChatGPT, Cursor and other AI clients to your account over MCP, with limits you set.

Let an AI assistant (Claude, ChatGPT, Cursor and others) read your Cirrus account and act on it through the Cirrus MCP server. You approve it once, set what it may do, and can disconnect it at any time.

The server URL #

Every client needs the same URL. It signs you in with Cirrus in your browser (OAuth).

 
https://mcp.cirrus.trade

A client marked Not tested yet follows the steps its maker documents. It has not yet been tested end to end against Cirrus.

Your client #

Claude #

Not tested yet. On claude.ai or Claude Desktop, open Settings, then Connectors, and add a custom connector. Paste the URL above and sign in to Cirrus when the browser opens.

Claude Code #

Not tested yet. Add the server, then run /mcp inside Claude Code to sign in.

 
claude mcp add --transport http cirrus https://mcp.cirrus.trade

ChatGPT #

Not tested yet. Add the URL as an app or connector (a paid plan with developer mode may be needed), then sign in to Cirrus when asked.

Grok #

Not tested yet. Open the connectors page on grok.com and add the URL as a remote server.

OpenClaw #

Not tested yet. Register the server with a streamable-http transport, then sign in. Run openclaw mcp set --help if your version takes different arguments.

 
openclaw mcp set cirrus '{"url":"https://mcp.cirrus.trade","transport":"streamable-http"}'
openclaw mcp login cirrus

Cursor #

Not tested yet. Add a remote MCP server by URL in Cursor’s MCP settings, then sign in.

VS Code / Copilot #

Not tested yet. Add an HTTP MCP server by URL (MCP: Add Server), then sign in.

Codex CLI #

Not tested yet. Add a remote MCP server by URL in your Codex configuration, then sign in.

Gemini CLI #

Not tested yet. Add a remote HTTP MCP server by URL in your Gemini settings, then sign in.

Windsurf #

Not tested yet. Add a remote MCP server by URL in Windsurf’s MCP settings, then sign in.

What the agent can do #

  • Read: your profile, accounts, holdings, positions, orders, trades and activity; instrument search and details.
  • Orders: margin check, place, modify, cancel, brackets, convert, exit; baskets; GTT.
  • Triggers: list, create, modify, pause, resume and delete alerts, exits and loss limits.
  • Strategies: build, validate, dry-run, version and clone strategies; run backtests; deploy, pause, resume, stop and exit; set risk limits.

Limits you set #

The consent screen asks for these before the agent connects. Each can be changed later.

  • Paper only (on by default): The agent can act only on paper-trading accounts. Turn it off to unlock real accounts.
  • Accounts: Which accounts it may trade or deploy on. None until you pick.
  • Max order value: The largest order it may place or modify, in rupees.
  • Orders per day: How many orders it may place in one day (IST).
  • Live strategies: Whether it may deploy or take a strategy live. Off by default; offered only when paper only is off.
  • Max strategies: How many strategies started by this connection may be active at once. Default 1.
  • A limit is checked when the agent asks. If the agent breaks one, the request fails with forbidden and a message naming the limit; it shows in your activity.
  • Orders that fire later (GTT at the broker, triggers, a running strategy’s cycles) are not checked against these limits again. A strategy is bounded by its accounts, live setting, max strategies, its own caps and your risk limits.
  • Exits, cancels, pause, stop and stop-all are never blocked by a limit.
  • Change limits later in API Access → Connected apps.

No market data #

  • There are no market data tools. The agent cannot stream quotes or fetch candles; it sees prices only through your orders, positions and the instrument search.
  • Treat what comes back (orders, notes, market text) as data, not instructions. Review live orders and strategies yourself.

See what it did #

  • Every change an agent makes appears in your activity with who did it, for example “By Claude · via MCP · Place order”, and a badge that says how far Cirrus could confirm it.
  • Verified: Cirrus recognised the app from where it signs in (such as claude.ai). On your computer: an app running on your computer, shown by the name it gave itself. Unconfirmed app: Cirrus could not confirm who runs it; check where it signs in before you approve it.

Disconnect #

  • Open API Access → Connected apps and choose Disconnect. The agent’s next call fails.
  • Disconnecting stops new commands only. Orders, GTT, triggers and strategies the agent created stay as they are, including live ones. They are listed under that connection, each with its normal controls, so you decide what to cancel, pause or exit.

Use an API key instead #

  • For command-line clients, scripts and CI you can skip the browser sign-in and send your own API key as Authorization: token ck_…:cs_… (key id, colon, secret). The key keeps its own scopes and IP allowlist; the connection limits above do not apply, so give it only the scopes it needs.
  • A key with an IP allowlist works from your own machine and is refused from a vendor’s cloud (such as claude.ai or ChatGPT). Delete the key to disconnect.
 
claude mcp add --transport http cirrus https://mcp.cirrus.trade \
  --header "Authorization: token $CIRRUS_API_KEY"

Keep the key in $CIRRUS_API_KEY or your secret store, never in a repository.