Brain Fight MCP

An Angel and a Devil, staged as a turn-by-turn debate over whatever you're stuck on — with two things most "AI debate" tools don't have: they remember how things have gone with you, and you can tell them how it actually turned out.

😇 Angel: Don't quit yet — line something up first.
😈 Devil: Quit. Today. Send the email.

Why this and not just asking Claude directly

Most multi-perspective decision tools are stateless: you ask, you get an analysis, it's forgotten. Brain Fight keeps a small relationship state (trust, annoyance, cooperation) per topic area that evolves across your actual conversations, and — if you come back later and mention what you did or how it went — it remembers that too. That memory doesn't just sit in a report: Devil might taunt you mid-debate with "last time you didn't listen to me and—", and the two of them notice their own history out loud the first time something rare happens — a freak run of agreement, one side winning five in a row, the tenth time you've hashed out a career decision together. That's the part nothing else here does.

It makes zero LLM calls of its own. The debate content is written by whichever model is calling it (Claude, in the normal case) — this server is just a fast, local, $0 state engine: it hands back structured performance instructions and tracks state in a local SQLite file.

Related MCP server: consensus-mcp

Install

Claude Desktop / Claude Code

Add to your MCP config (claude_desktop_config.json or equivalent):

{
  "mcpServers": {
    "brain-fight": {
      "command": "npx",
      "args": ["-y", "brain-fight-mcp"]
    }
  }
}

Restart your client. No API keys, no accounts, no network calls required.

Run it yourself

npx brain-fight-mcp            # stdio transport (default)
npx brain-fight-mcp --http     # Streamable HTTP on 127.0.0.1:8000

Tools

You generally won't call these by name yourself — you just tell Claude what you're stuck on, and it decides when to reach for these based on the server's built-in instructions (things like "should I..." or "I can't decide..." trigger it automatically; venting like "I'm so stressed" gets offered as an option rather than sprung on you).

Topic domains

Relationship trust and track record are bucketed by life area, so a mishandled snack decision doesn't dilute the trust built up around a real career call. The buckets: career, money, relationships, health, general (default). There's no keyword classifier picking this for you — the model driving the conversation (Claude) chooses the domain based on what you're actually asking about when it calls start_debate. You don't have to specify it yourself, but if it ever picks the wrong bucket, just say so and ask it to use a different one.

Track record & milestones

Two things carry across debates within a domain, and both are used sparingly by design — the instructions given to the model explicitly say "most turns should NOT mention this," because a callback that fires constantly stops feeling like memory and starts feeling like a stat sheet:

• Track record. If you've told it (via record_decision_outcome) how past decisions actually turned out, Angel or Devil may — rarely, when it genuinely fits — bring that up mid-debate ("last time you didn't listen to me and—"), not just in a closing summary. When you left a specific note on an outcome, it can reference that directly ("you said you bought the laptop on impulse and returned it a week later") instead of just a tally. Older outcomes fade rather than vanish: a regret from a year ago barely moves things, one from last week moves them a lot, and the framing is honest about how long ago it was ("a while back," not "last time," for anything more than a few months old).

• Milestones. A handful of one-time narrative beats, each fired at most once ever per domain: cooperation crossing an unusually high level for the first time, either side winning five debates in a row, or hitting the 10th debate in a domain. When one fires, Angel and Devil briefly react to it in character — genuinely surprised, not a scripted congratulations message — right before you get the actionable next step.

None of this is configurable per-milestone right now; it's meant to be a small set of rare, earned moments rather than a feature you tune.

Data & privacy

Everything is stored locally in a single SQLite file — nothing is sent anywhere except to the model that's already driving your conversation.

• Default location: ~/.brain-fight-mcp/state.sqlite3

• Override with BRAIN_FIGHT_DB_PATH=/path/to/file.sqlite3

• Wipe it any time with reset_relationship (one domain or everything — this also resets any milestones reached, so they can genuinely happen again from scratch) or clear_database (everything, all sessions)

• Completed debate transcripts are auto-pruned after 30 days by default (the durable summary — context, positions, winner — is kept indefinitely; only the raw turn-by-turn transcript is pruned). Adjust with BRAIN_FIGHT_RETENTION_DAYS=<n>, or set it very high to keep everything.

Running over HTTP

For remote/web clients instead of local stdio:

npx brain-fight-mcp --http --port 3000 --token <secret>

If you bind to 0.0.0.0 (or ::) without a --token, the server logs a loud warning and starts anyway — anyone who can reach the port can call your tools. Set a token before exposing this beyond your own machine.

Health check: GET /health. MCP endpoint: ALL /mcp.

Development

npm install
npm run build       # tsup → dist/
npm test            # vitest

These assume standard build/test script names in package.json. Adjust if yours are named differently — this README was written without seeing your actual package.json, since it wasn't part of what you shared with me.

License

(add your license here)