MCP Integration

Vāda MCP

Vāda exposes deliberation as MCP tools (vada__consult and vada__deliberate). Any MCP-compatible client can call them. The hosted server runs at https://vada.attalabs.dev/api/mcp no local install required. Authentication via Vāda API key. Provider model calls run on your configured BYOK keys.

Getting started

Connect to the hosted server — no local setup required.

Claude.ai's MCP config UI lets you enter a URL and custom headers. Open Settings → Integrations → MCP and add a new server with the following config.(Exact UI flow TBD — Principal to verify)

{
  "url": "https://vada.attalabs.dev/api/mcp",
  "headers": {
    "Authorization": "Bearer vada_pk_your_key_here"
  }
}

API Key

Generate a Vāda API key in Settings → MCP. Paste it as the Bearer token in your client config. Treat it like a password— it gives full access to your account's deliberation tools.

1

Go to Settings → MCP and generate a new API key.

2

Paste the key as the Authorization: Bearer value in your client config above.

Provider Keys

The hosted server runs deliberations on your provider keys. Configure them in Settings → API Keys. Keys are encrypted at rest using server-managed keys.

Unlike the web app's browser-only BYOK, this is a different trust model: Vāda's server decrypts your keys to make provider calls on your behalf. See the trust model for more detail.

Tools

vada__consult

Single-shot reviewer chain. Pass a question and 2–5 reviewer roles. Each reviewer responds independently.

Use vada__consult to review this proposal: [your question here]
Reviewers: critic, strategist, devils_advocate
vada__deliberate

Multi-round team deliberation. Pass a question and optionally a team name. Returns a synthesized conclusion. Default team: sparring.

Use vada__deliberate to analyze: [your question here]
Team: war-room

Teams

Pass one of these team names to vada__deliberate to select the deliberation format.

sparring

Sparring

Two agents go head-to-head across three rounds. Strategist and Critic stress-test each other's reasoning before a synthesizer reconciles.

crucible

Crucible

Four agents debate across three rounds. Strategist proposes, Critic challenges, Devil's Advocate breaks, Synthesizer reconciles. Two auditors verify before delivery.

war-room

War Room

Six agents debate across three rounds — strategy, criticism, adversarial framing, synthesis, research, and execution. For high-stakes decisions.

vada-reviewers

Reviewers

Three different LLMs review your problem in parallel — Gemini, GPT, and Grok give independent perspectives. Vendor diversity over role specialization.

vada-reviewers-synthesis

Reviewers + Synthesis

Three different LLMs review independently, then a synthesizer extracts consensus, contradictions, and unique insights into a structured analysis.

Alternative

Local stdio server

For local development or self-hosted setups, you can run the MCP server as a local process. Clone the repo, build the server, and point your client at the built binary.

1

Prerequisites

  • Node.js 18+ (nodejs.org)
  • Anthropic API key from console.anthropic.com
  • Postgres database URL (optional — omit to skip session persistence)
  • Clone the repo: git clone https://github.com/attaai/attaai
  • Build the MCP server: cd packages/mcp-server && bun run build
2

Config file location

~/Library/Application Support/Claude/claude_desktop_config.json

3

Add to config

{
  "mcpServers": {
    "vada": {
      "command": "node",
      "args": ["/path/to/attaai/packages/mcp-server/dist/index.js"],
      "env": {
        "ANTHROPIC_API_KEY": "sk-ant-...",
        "DATABASE_URL": "postgresql://...",
        "VADA_USER_ID": "your-clerk-user-id"
      }
    }
  }
}

Set VADA_USER_ID to your Vāda account ID to see consultations in the dashboard. Find it in Settings.

4

Restart

Quit Claude Desktop completely and relaunch. The MCP server starts automatically on next launch.

5

Verify

Open a new conversation and type "/" — you should see vada__consult and vada__deliberate in the tool list.

Troubleshooting

Tool not appearing in client

Check that your API key is valid and that your client can reach vada.attalabs.dev. Some clients require a restart after config changes.

401 Unauthorized

API key is invalid or has been revoked. Regenerate it in Settings → MCP.

400 "missing provider key"

The hosted server needs a provider key to make model calls. Configure one in Settings → API Keys.

Request times out

Deliberation runs can take 3–6 minutes depending on team size and model. Increase your client timeout if it cuts off before the result arrives.