Run the Council from inside Claude Desktop, Cursor & Windsurf

Council AI's MCP server (Ultra tier, $99.99/mo) exposes nine tools to any MCP-aware client: council_query (fan a prompt across N frontier models, return moderator-synthesized consensus + agreement score, key disagreement, and an optional devil's-advocate synthesis mode), council_review (multi-model code review of a diff — verdict-first: SHIP/NO-SHIP, findings confirmed by 2+ models, attributed single-model dissents), council_query_with_rag (same as council_query plus retrieval from your personal Council RAG library — unique to Council), library_search, library_list, library_upload, library_delete, council_models, and council_usage. The endpoint is https://mcp.council-ai.app/mcp, authenticated via per-user Personal Access Tokens you generate in Settings. Works with Claude Desktop, Cursor, Windsurf, Claude Code, ChatGPT Connectors, and any other MCP-aware client. Setup takes about 60 seconds.

What the MCP server exposes

• council_query — fan a prompt across N frontier models in parallel. Returns the moderator's synthesized answer plus an agreement score across the council.
• council_review — multi-model code review. Send a diff (plus an optional focus: bugs, security, performance, architecture) and get a verdict-first synthesis: SHIP/NO-SHIP, findings confirmed by 2+ models, and single-model dissents attributed with reasoning. A free Claude Code skill wraps it: github.com/omarkeshk/council-review.
• council_query_with_rag — same as council_query, but each model first retrieves relevant chunks from your personal RAG library. Unique to Council; no other multi-LLM MCP bundles this.
• library_search — semantic search over the user's RAG library.
• library_list — enumerate documents in the user's RAG library.
• library_upload — add a document to your RAG library from inside your AI client.
• library_delete — permanently remove a document, its chunks, and its file.
• council_models — list AI models available to the user's tier.
• council_usage — current monthly cost-budget consumption.

Setup in Claude Code

One command in your terminal (--scope user makes it available in every project):

claude mcp add --transport http --scope user council-ai \
  https://mcp.council-ai.app/mcp \
  --header "Authorization: Bearer csa_YOUR_PAT_HERE"

Generate your PAT in Settings → MCP tokens.

Setup in Claude Desktop

Claude Desktop launches local MCP servers only, so the config bridges to our remote endpoint through mcp-remote. Open claude_desktop_config.json and add:

{
  "mcpServers": {
    "council-ai": {
      "command": "npx",
      "args": [
        "-y",
        "mcp-remote",
        "https://mcp.council-ai.app/mcp",
        "--header",
        "Authorization:${COUNCIL_AUTH}"
      ],
      "env": {
        "COUNCIL_AUTH": "Bearer csa_YOUR_PAT_HERE"
      }
    }
  }
}

The Authorization:${COUNCIL_AUTH} form (no space, env-interpolated) is deliberate — it also works on Windows, where spaces inside args break. Restart Claude Desktop; the council tools appear in the tools menu.

Setup in Cursor

Add to ~/.cursor/mcp.json (or Cursor Settings → MCP):

{
  "mcpServers": {
    "council-ai": {
      "url": "https://mcp.council-ai.app/mcp",
      "headers": { "Authorization": "Bearer csa_YOUR_PAT_HERE" }
    }
  }
}

Cursor calls the council tools from its agent loop the moment they're listed.

Setup in Windsurf

Windsurf reads ~/.codeium/windsurf/mcp_config.json and names the key serverUrl (not url):

{
  "mcpServers": {
    "council-ai": {
      "serverUrl": "https://mcp.council-ai.app/mcp",
      "headers": { "Authorization": "Bearer csa_YOUR_PAT_HERE" }
    }
  }
}

Why ship a multi-LLM MCP at all

Single-model coding assistants are great until they're confidently wrong. The MCP server gives your agent loop a "phone a friend" tool — when Claude inside Cursor is uncertain about a refactor, it calls council_query and gets verification from GPT-5.5, Gemini 3 Pro, and DeepSeek V4 in one round-trip — and before you merge, council_review has models from multiple labs review the same diff and returns a SHIP/NO-SHIP verdict with confirmed findings and dissents. With council_query_with_rag, those models also see your codebase notes and architecture docs from your RAG library.

Frequently asked questions