collider.runMulti-agent AI debate

Collider in your AI assistant

Collider is an MCP server. Add it to Claude Code, Claude Desktop, Cursor, VS Code or Gemini CLI, and your assistant can start a refereed debate between rival frontier models, wait for the verdict, and use it mid-task. It can also run a whole prompt set through the room as a bake-off. No SDK, no glue code.

MCP endpointhttps://collider.run/mcp

How a collision works

Your assistant calls start_collision with a brief: the question, claim or document to examine. Collider seats rival models (ChatGPT, Gemini, Claude, Grok, Mistral) in one room, runs the debate server-side, and computes a verdict. A collision typically takes 5 to 15 minutes, so start_collision returns a collisionId within seconds and the assistant polls get_collision until it completes.

The result is never a black-box score. The verdict comes with the full transcript: every turn, every model, every assessment, so any conclusion can be traced to the exchange that produced it.

1 · Get an API key

  1. Sign in at collider.run.
  2. Open your Account and generate an API key under API Key.
  3. Copy the clsk_… value immediately. It is shown once; the server keeps only a hash.

Calls are billed from the same credit pool as your web account, at the menu prices below. Treat the key like a password: it can start paid work. Revoke or regenerate it in your Account at any time.

2 · Connect your client

Claude Code

claude mcp add --transport http collider https://collider.run/mcp \
  --header "Authorization: Bearer clsk_YOUR_KEY"

Cursor

~/.cursor/mcp.json
{
  "mcpServers": {
    "collider": {
      "url": "https://collider.run/mcp",
      "headers": { "Authorization": "Bearer clsk_YOUR_KEY" }
    }
  }
}

VS Code

.vscode/mcp.json
{
  "servers": {
    "collider": {
      "type": "http",
      "url": "https://collider.run/mcp",
      "headers": { "Authorization": "Bearer clsk_YOUR_KEY" }
    }
  }
}

Claude Desktop

Claude Desktop reaches remote servers through the mcp-remote bridge (needs Node.js installed).

claude_desktop_config.json → mcpServers
{
  "mcpServers": {
    "collider": {
      "command": "npx",
      "args": [
        "-y", "mcp-remote", "https://collider.run/mcp",
        "--header", "Authorization: Bearer clsk_YOUR_KEY"
      ]
    }
  }
}

Gemini CLI

~/.gemini/settings.json
{
  "mcpServers": {
    "collider": {
      "httpUrl": "https://collider.run/mcp",
      "headers": { "Authorization": "Bearer clsk_YOUR_KEY" }
    }
  }
}

Any other MCP client works too, as long as it speaks the Streamable HTTP transport and can send an Authorization header. The web connectors of ChatGPT and claude.ai require OAuth sign-in, which this server does not offer yet.

3 · Ask for a collision

Once connected, plain language is enough:

“Run a collision on the draft below and give me the referee verdict.”

“Start a review-depth collision with web grounding on this claim: …
 Poll until it completes, then summarise where the models disagreed.”

“Bake off these 12 prompts with their expected answers across the seats,
 then tell me where the room agreed and was still wrong.”

The assistant fills in the tool calls: it starts the run, polls every 30 to 60 seconds (every few minutes for a bake-off), and reads the result when the room is done.

Tools

start_collision

ParameterTypeDefaultNotes
briefstringrequiredThe question, claim or document. Self-contained, max 40,000 characters. The debaters see nothing else.
roundsinteger 1–42Each round beyond 2 adds 50% to the price.
personaenumresearchersresearchers · panel · friends · comedians
responseLengthenumstandardbrief · standard · detailed · review (per-turn depth; sets the base price)
webGroundingbooleanfalseProvider-native web search per seat, +50% price.
lineupstring[]active lineup2–5 of chatgpt · gemini · claude · grok · mistral
modelsobjectactive buildsPer-seat model override; a seat only accepts names from its own vendor family.
tagstringrunLabel for your own bookkeeping.

Returns { collisionId, status: "queued", credits: { charged, breakdown } }.

get_collision

ParameterTypeDefaultNotes
collisionIdstringrequiredFrom start_collision.
include_transcriptbooleanfalseThe full trace can be very long; it is withheld unless asked for.

While running: progress as turnsCompleted / totalTurns. When complete: the verdict, per-seat usage, and (on request) the transcript.

credit_balance

No parameters. Returns the balance of the account behind the key, plus the current price menu.

start_bakeoff

ParameterTypeDefaultNotes
title, audiencestringrequiredPrinted on the bake-off report; the audience shapes the brief.
itemsobject[]required1–40 of { id?, prompt, context?, expected?, rubric? }. expected and rubric are grading-only, never shown to a seat.
roundsinteger 2–32Round 1 is blind; later rounds cross-examine.
responseLengthenumstandardAs for collisions.
webGroundingbooleanfalseSeats without a web tool are told they have none.
lineupstring[]active + custom2–5 seat keys, registry and/or custom1..custom4.
customSeatsobject[]noneYour own OpenAI-compatible endpoints as seats; see bake-off docs.
notifyEmail, notifystring, booleanaccount, trueWhere the “ready” email goes; notify:false for none.

Returns { reportId, status: "queued", cellsTotal, credits: { charged, breakdown } }. Full field reference on the bake-off page.

get_report

ParameterTypeDefaultNotes
reportIdstringrequiredFrom start_bakeoff.
include_fullbooleanfalsePer-item detail is summarised unless asked for.

While running: cellsComplete / cellsTotal. When complete: headline counters, per-seat table, per-item results, and links to the PDF and the signed-in report page.

Pricing and limits

  • Menu per collision (two rounds): standard 50 · detailed 100 · review 300 credits. Web grounding +50%; each round beyond 2 adds +50%; 50-credit minimum.
  • Bake-off: per item plus a synthesis fee, printed in the accept response before anything runs (see bake-off pricing).
  • The price is charged when the run is accepted and printed in the response, so the charge is never a surprise. A run that fails or is interrupted is refunded automatically; a bake-off refunds the items that did not run.
  • One collision or report runs at a time per account, with a daily cap shown in your Account.
  • Credits are one pool across the web app and the API. Top up at collider.run/pricing.

Prefer raw HTTP?

The same key drives the HTTP API directly:

curl -X POST https://collider.run/api/collide \
  -H "Authorization: Bearer clsk_YOUR_KEY" \
  -H "Content-Type: application/json" \
  -d '{"brief": "Claim plus full context here.", "responseLength": "standard"}'

curl https://collider.run/api/collide/COLLISION_ID \
  -H "Authorization: Bearer clsk_YOUR_KEY"

Questions

Use the feedback button inside the app; it reaches the founders directly.

Measurement method and validation published (pre-registered study, report and data on OSF); calibration corpus and prompts are ours.

collider
PricingAPITermsPrivacy

© 2025–2026 Collider. All rights reserved.