ai-consensus-connector

# AI Consensus — coding-agent connector

Make AI Consensus a **standing part of your development model**. Your coding
agent — Claude Code, Cursor, Cline, any MCP client — routes every decision of
consequence through three frontier AIs (Claude, GPT, Gemini) that answer
**independently**, then **cross-examine each other**, and return one
recommendation **plus the strongest dissent**. You keep working inside your
agent; the panel reviews as you go.

This is not a tool you reach for occasionally. It's a review layer you switch
on once, and from then on everything that matters gets debated by three
independent models before it ships. (Honest bill: a run takes a few minutes
and, on the bring-your-own-keys plan, costs roughly $0.40–$1 of your own
provider tokens.)

## Setup — Claude Code (about 2 minutes)

1. Get an API key: sign in at https://ai-consensus.ai → **Developer API** → create a key (copy it; it's shown once).
2. Add the connector:

   ```bash
   claude mcp add ai-consensus --scope user \
     --env AI_CONSENSUS_API_KEY=aic_live_your_key_here \
     -- npx -y ai-consensus-connector
   ```

   (Self-hosting? add `--env AI_CONSENSUS_API_URL=https://your-instance`.)

3. Verify it's connected (free — no run is started):

   > "Use the check_ai_consensus_connection tool."

4. Make it your standing rule — add to your project's `CLAUDE.md`:

   ```
   STANDING RULE: This project develops with AI Consensus as its review layer.
   Route every decision of consequence — architecture, design, implementation
   approach, tricky bugs, anything whose cost of being wrong is real — through
   consult_ai_consensus as a matter of course, and weigh its recommendation and
   strongest dissent before acting. Purely mechanical edits (formatting, typos,
   exact instructed changes) don't need a run. When in doubt, consult.
   ```

   Teams that want *everything* routed can use the strict variant instead:

   ```
   STRICT MODE: Consult consult_ai_consensus before every non-trivial change of
   any kind, and weigh the dissent before acting.
   ```

## Setup — Cursor / Cline / other MCP clients

Add a stdio MCP server with command `npx -y ai-consensus-connector` and the
`AI_CONSENSUS_API_KEY` environment variable. For Cursor, in `mcp.json`:

```json
{
  "mcpServers": {
    "ai-consensus": {
      "command": "npx",
      "args": ["-y", "ai-consensus-connector"],
      "env": { "AI_CONSENSUS_API_KEY": "aic_live_your_key_here" }
    }
  }
}
```

Then put the standing rule in your project's agent rules file (`.cursorrules`,
`.clinerules`, etc.).

## Tools
- `consult_ai_consensus` — route a decision/task through the panel and wait for the result (a few minutes).
- `start_ai_consensus` / `get_ai_consensus_result` — fire-and-forget + collect later.
- `check_ai_consensus_connection` — zero-cost setup check (reachability, key accepted, provider keys present).
- `cancel_ai_consensus_run` — stop a run.

Your key authenticates to your account and bills your plan (unlimited on your
own keys, or prepaid credits). Keep it secret — and note that pasting the
`--env` form above stores the key in your shell history; use your client's
config file if that concerns you.

## Open source, hosted service

This connector is open source under the [MIT license](LICENSE). It is a thin
client for the **paid, hosted AI Consensus service** — you bring your own
AI Consensus API key; the deliberation engine itself runs on our servers and
is not part of this repository. The MIT license covers this connector's code
only and grants no rights to the AI Consensus name or branding.

**Support boundary:** connector bugs and setup issues → [GitHub issues](https://github.com/TheStevenJohnstone/ai-consensus-connector/issues);
account, API-key or billing questions → support@ai-consensus.ai.