The number one piece of feedback we got after launching TRACE was this: "It's useful, but I keep forgetting to run the commands."

Which is ironic. TRACE exists because AI tools forget to follow process. And the solution to AI forgetting process was... a tool that humans had to remember to run.

So we fixed it. TRACE v1.0.2 ships with an MCP server that makes the entire system autonomous. Your AI tool calls TRACE automatically during conversations. You never type trace gate start again.

What changed

Before: TRACE was reactive. You coded, ran trace gate end, and found out what broke. The damage was already done. You were discovering problems, not preventing them.

After: TRACE is proactive. Your AI tool reads the project state before starting, checks blast radius before modifying files, validates coherence after changes, and checks dependency policy before adding packages. All automatically.

The difference is like the difference between a smoke alarm and a fire prevention system. Both are useful. One goes off after the fire starts. The other prevents ignition.

How it works

MCP (Model Context Protocol) is how AI tools connect to external services. Claude Code, Cursor, and Kiro all support it. TRACE exposes 6 tools through an MCP server:

trace_context — The AI calls this at the start of every conversation. It returns the full project state: anchors, consumers, debt, plan items. The AI reads this to understand the structural landscape before writing a single line.

trace_impact — The AI calls this before modifying any file that might be an anchor. It returns every consumer that depends on that file. The AI now knows: "If I change user.ts, I must also update auth.ts, cart.ts, and docs/users.md." It doesn't need to be told. It knows.

trace_check — The AI calls this after making changes. One line if everything passes. Full details if something is wrong. The AI fixes issues before presenting results to you.

trace_deps_check — The AI calls this before adding a new package. "Is lodash allowed? Does it pass our license policy? Is it on the blocked list?" The AI gets an answer before running npm install.

trace_log — The AI calls this when done. Records what happened to PROJECT_LOG and HANDOFF so the next session has full context.

trace_status — Quick health check. Anchor count, consumer count, debt level. Lightweight.

Setup takes 2 minutes

Install TRACE globally if you haven't:

npm install -g trace-coherence

For Claude Code, add to your .claude.json:

{
  "mcpServers": {
    "trace": {
      "command": "trace-mcp"
    }
  }
}

For Cursor, same config in .cursor/mcp.json. For Kiro, add trace-mcp as an MCP server in .kiro/settings/mcp.json.

Restart your AI tool. TRACE tools appear automatically. Done.

Token overhead is minimal

The obvious concern: "Won't the MCP calls eat up my token budget?"

We built four optimizations specifically for this:

Total per session: about 1,300 to 2,100 tokens. For comparison, one undetected drift incident costs 30,000 to 50,000 tokens of debugging.

What this means for your workflow

Your daily workflow changes from this:

# Old way (manual)
trace gate start
# ... code ...
trace impact user_model
# ... more code ...
trace gate end

To this:

# New way (autonomous)
# Just code. TRACE handles itself.

The AI reads context automatically. Checks impact automatically. Validates automatically. Logs automatically. You focus on the actual work.

The pre-commit hook is still there as the final safety net. If somehow the AI skips everything, the commit is blocked. But with the MCP server active, that almost never happens.

This works with any AI tool

TRACE is not tied to Claude Code. The MCP server works with any tool that implements the Model Context Protocol: Claude Code, Cursor, Kiro, and any future tool that supports MCP. Your verification layer is portable across AI tools, even if your team uses different ones.

That's the point. The steering is tool-specific (CLAUDE.md, .cursor/rules, Kiro steering). The verification is universal.