AI Integration

Tracey exposes its coverage analysis as MCP (Model Context Protocol) tools, letting AI assistants like Claude Code query requirements, find uncovered code, and add annotations.

Setup

Option 1: MCP config

Add a .mcp.json file to your project root:

json

{
    "mcpServers": {
        "tracey": {
            "command": "tracey",
            "args": ["mcp"]
        }
    }
}

The MCP server auto-starts the tracey daemon when it connects.

Option 2: Install the skill

bash

tracey skill install

This installs a skill file that teaches AI assistants how to use tracey's tools effectively. By default it installs for both Claude Code and Codex. Use --claude or --codex to install for just one.

Available tools

Tool	Purpose
`tracey_status`	Coverage overview — shows configured specs, prefixes, and percentages
`tracey_uncovered`	Requirements without `impl` references
`tracey_untested`	Requirements without `verify` references
`tracey_stale`	References pointing to older rule versions
`tracey_unmapped`	Source tree with coverage — shows code without requirement references
`tracey_rule`	Full details about a specific requirement
`tracey_config`	Display current configuration
`tracey_validate`	Check for broken references, naming issues, duplicates
`tracey_reload`	Reload config and rebuild data
`tracey_config_include`	Add an include pattern to an implementation
`tracey_config_exclude`	Add an exclude pattern to an implementation

Filtering

tracey_uncovered, tracey_untested, and tracey_stale accept optional parameters:

spec_impl — filter to a specific spec/implementation (e.g., "myapp/rust")
prefix — filter by requirement ID prefix (e.g., "auth." to see only auth requirements)

tracey_unmapped accepts an optional path parameter to zoom into a directory or file.

Workflow

A typical session with an AI assistant:

Check coverage — the assistant calls tracey_status to see what specs exist, what prefix to use, and current coverage percentages.
Find work — tracey_uncovered shows requirements that lack implementations. tracey_untested shows requirements without tests.
Read requirements — tracey_rule fetches the full text of a specific requirement, along with all existing references.
Annotate code — the assistant adds r[impl req.id] or r[verify req.id] comments to the appropriate code.
Verify — tracey_status again confirms coverage improved.

Example conversation

You: "Add spec annotations to the auth module"

The assistant calls tracey_status, sees prefix r and 45% coverage, then calls tracey_uncovered to find auth.login, auth.token-expiry, and auth.rate-limiting are uncovered. It reads each requirement with tracey_rule, finds the implementing functions, and adds the annotations.

Response format

MCP tool responses include:

Status header — current coverage for all spec/implementation pairs
Delta — what changed since the last query (newly covered requirements)
Hints — suggestions for what to query next

Responses are formatted as human-readable text, not JSON.

Single vs. multiple specs

When only one spec and one implementation are configured, tools use them by default — no need to specify spec_impl. When multiple exist, tools either auto-detect from context or ask you to specify.