DocumentationOperate

Working with agents

Updated: 2026-08-09

canary mcp is a deliberately small, read-only adapter for local AI clients. It reads the same typed daemon authority as the CLI and paired app, but it has no streaming resources and cannot preview, place, modify, cancel, exercise, or transmit a broker action.

For exact parameters, use the generated MCP tools reference.

Setup

The Claude Code plugin carries the Canary skill, safety hooks, and MCP server configuration. Install the canary binary separately, then verify the host:

claude plugin details canary@canary
claude mcp list
canary status

Fully quit and relaunch the host after updating Canary. MCP hosts keep their stdio child alive across chats; canary restart restarts the shared daemon, not that child process.

Start with the brief

For “what matters today?” or “what changed since the close?”, use canary_brief. It is the same combined post-trade and pre-trade report used by the CLI and paired app. It includes current market-risk posture, portfolio and PnL context, rulebook adherence, actions, and explicit missing or stale inputs. The MCP read never stamps or acknowledges the report.

Drill into narrower evidence only when needed:

Monitor profile

canary mcp --profile monitor

The low-token monitor profile exposes only canary_brief and canary_status. Use the brief for the desk decision surface and status only to diagnose connectivity or degraded inputs.

Safety and evidence

MCP output is evidence, never broker-write authority. Missing or stale data is not a clean result, and a cached or held market-risk value is display context only. Multi-account responses must identify one current account and mode; otherwise the daemon refuses account-scoped authority.

Protection and exercise candidates are discovery records. Actual submission requires the separate gated CLI or paired-app flow, a fresh exact preview or preflight, and an explicit transaction-specific instruction from the user.

Reference