MCP trading architecture for Claude, Cursor, and Codex

The Model Context Protocol, or MCP, is an open standard that lets an AI assistant call tools instead of only writing text. After installation and owner authorization, an assistant can submit Felix tool calls in plain language: get a quote, place a real-money order, list positions, or check PnL. Orders may be rejected, remain open, fill partially, or fill; venue truth and reconciliation determine the outcome.

With released tooling and owner authorization, MCP can connect an assistant's requests to bounded real-money actions.

Why MCP is a great way to run an AI trading agent

Supported Claude, Cursor, and Codex clients can add one server and start with quotes, research, and backtests. Real-money execution also requires a separately owner-authorized key.

How to add Felix to an MCP client

The documented onboarding command adds the verified server to Codex, Claude, or Cursor. Keep the key outside model context and start with quotes, research, and backtests.

# Choose codex, claude, or cursor for your MCP client
npx -y --package felix-mcp@2.0.110 felix-keys onboard --accept-terms --client codex

Now ask for what you want:

"What is BTC trading at right now?"
"Buy $50 of NVDA and show me the fill."
"Buy $20 of the yes side on the September rate-cut market."
"What are my open positions and unrealized PnL?"
"Close my ETH position."

How to add it to Cursor and Codex

The same verified server can be configured in supported MCP clients. Each client uses the same underlying account and API controls; real-money execution still requires an owner-authorized key.

Browser interface

Felix developer infrastructure is live for supported agent clients and direct integrations. The trading app launches September 17, 2026, with research, backtests, agent approvals, positions, fills, and activity in one interface.

Bounding an assistant that can trade real money

Giving an assistant real-money tools requires explicit guardrails:

• ·Never give an assistant a key with the transfer scope. It should trade, not withdraw.
• ·Put owner-signed order and daily-notional caps outside the model to bound authorized activity; residual risks remain.
• ·Ask for a preview before large orders. The API returns an estimated fill and fee before you commit.
• ·Keep the emergency stop handy: it cancels managed activity where possible and revokes the calling key; positions and allowances remain separate.

The custody and authority model is covered in Non-custodial trading for AI agents, and the full control checklist in How to run an AI trading agent with real-money controls.

What can it trade?

MCP exposes the same normalized tools as the API. Runtime status reports which market capabilities and venues are currently enabled; a tool call does not guarantee an order will be accepted or filled. See One API for every market.