Anchord MCP Server

Identity resolution and pre-write safety checks for AI agents.

An MCP server that gives AI agents access to the Anchord identity resolution API. Resolve companies and people to canonical AnchorIDs, run pre-write safety checks, and export golden records — through the standard MCP tool interface.

Hosted API-backed. This MCP server is a thin proxy to the Anchord SaaS platform. All scoring, matching, validation, and data persistence happen server-side. No business logic runs locally.

Read-only by design. Anchord never writes to your external systems (CRMs, databases, etc.). guard_write evaluates a proposed write and returns allowed/blocked — the caller decides whether to proceed.

Quick start

1. Get an API key

Sign up at app.anchord.ai/signup and create an API key in Settings > API Keys.

2. Run with npx (no install)

ANCHORD_API_KEY=<YOUR_ANCHORD_API_KEY> npx -y @anchord/mcp-server

That's it. The server starts over stdio and is ready for MCP clients.

3. Or connect to the hosted remote (zero install)

No local install needed. Point any MCP client that supports remote HTTP transport at the hosted endpoint:

{
  "mcpServers": {
    "anchord": {
      "url": "https://mcp.anchord.ai/mcp",
      "headers": {
        "Authorization": "Bearer <YOUR_ANCHORD_API_KEY>"
      }
    }
  }
}

See docs/remote.md for full details, client compatibility notes, and a local fallback if your client does not yet support remote MCP.

Related MCP server: io.magellandata/spotlight

MCP client setup

Cursor (local stdio)

Add to .cursor/mcp.json (workspace) or ~/.cursor/mcp.json (global):

{
  "mcpServers": {
    "anchord": {
      "command": "npx",
      "args": ["-y", "@anchord/mcp-server"],
      "env": {
        "ANCHORD_API_KEY": "<YOUR_ANCHORD_API_KEY>"
      }
    }
  }
}

See examples/cursor-mcp.json.

Claude Desktop

Add to your Claude Desktop config (~/Library/Application Support/Claude/claude_desktop_config.json on macOS, %APPDATA%\Claude\claude_desktop_config.json on Windows):

{
  "mcpServers": {
    "anchord": {
      "command": "npx",
      "args": ["-y", "@anchord/mcp-server"],
      "env": {
        "ANCHORD_API_KEY": "<YOUR_ANCHORD_API_KEY>"
      }
    }
  }
}

See examples/claude-desktop-config.json.

Remote MCP (for clients that support HTTP transport)

For zero-install remote access, use the hosted endpoint instead of a local stdio process. This works with any MCP client that supports the url + headers configuration format:

{
  "mcpServers": {
    "anchord": {
      "url": "https://mcp.anchord.ai/mcp",
      "headers": {
        "Authorization": "Bearer <YOUR_ANCHORD_API_KEY>"
      }
    }
  }
}

No Node.js, no npx, no Docker required. If your client does not yet support remote MCP, use the local stdio setup above. See docs/remote.md for full details.

Docker

docker build -t anchord-mcp .
echo '{"jsonrpc":"2.0","id":1,"method":"initialize",...}' | \
  docker run --rm -i -e ANCHORD_API_KEY=<YOUR_ANCHORD_API_KEY> anchord-mcp

Or use the compose file for local testing:

cp examples/env.example .env
# Edit .env with your API key
docker compose up

Environment variables

See docs/auth.md for details on authentication and tenant scoping.

Available tools

Full parameter reference: docs/tools.md

Safe agent workflow

The recommended sequence for agents writing to external systems:

1. ingest_record        Push the source record into Anchord
                        (optional if using OAuth integrations)

2. resolve_company      Match to a canonical AnchorID
   or resolve_person    → status: resolved | not_found | needs_review

3. IF needs_review      STOP. Do not write.
                        Surface candidates to the user.
                        Direct them to the Review Queue.

4. guard_write          Evaluate the proposed write
                        → allowed: true | false (with block codes)

5. IF allowed           The agent performs the external write.
                        Anchord never writes.

6. Log request_id       Every response includes a request_id
                        for audit trail and debugging.

Use get_entity or get_entity_export at any point to inspect AnchorID details or retrieve the merged golden record.

Handling needs_review

Only resolve_* returns needs_review. It means Anchord found multiple plausible matches and cannot auto-resolve with confidence.

For agents:

1. Do not write. The data is ambiguous.

2. Surface the candidates to the user — the response includes entity IDs and match scores.

3. Direct the user to the Review Queue: https://app.anchord.ai/app/queues/needs-review

4. Retry later. Once a human resolves the ambiguity, subsequent resolve calls return resolved.

Example agent message:

I tried to resolve "Acme Corp" but Anchord found multiple possible matches. A human needs to review this in the Review Queue. I'll retry after it's resolved.

Error handling

When the API returns 4xx/5xx, the MCP tool response is marked isError: true with a structured payload:

{
  "error": "[422] BATCH_TOO_LARGE: Batch size must not exceed 100 records. (request_id: req_01ABC123)",
  "status_code": 422,
  "request_id": "req_01ABC123",
  "details": { "records": ["Too many records."] }
}

• request_id is always present — from the API response body, x-request-id header, or a client-generated UUID.

• details contains validation errors when available (null for non-JSON errors).

• API keys are never included in error messages.

Architecture

Local (stdio)

MCP Client (Cursor / Claude Desktop / etc.)
    │  stdio (JSON-RPC)
    ▼
┌──────────────┐
│  MCP Server  │  Node.js + TypeScript
│  (this pkg)  │  Zod schemas · no business logic
└──────┬───────┘
       │  HTTPS + Bearer auth
       ▼
┌──────────────┐
│  Anchord API │  Hosted SaaS — scoring, matching,
│              │  persistence, tenant isolation
└──────────────┘

Hosted remote (HTTP)

MCP Client
    │  HTTPS POST + Bearer token
    ▼
┌────────────────────────┐
│  mcp.anchord.ai        │  CloudFront (TLS, routing)
└───────────┬────────────┘
            ▼
┌────────────────────────┐
│  Lambda (stateless)    │  Per-request MCP server
│  Bearer → ApiClient    │  No stored secrets
└───────────┬────────────┘
            │  HTTPS + Bearer auth
            ▼
┌────────────────────────┐
│  Anchord API           │  Same hosted SaaS backend
└────────────────────────┘

Both paths expose the same 11 MCP tools and connect to the same API.

FAQ

Is Anchord self-hosted?

No. Anchord is a hosted SaaS platform. This MCP server is a thin client that calls the Anchord API. You need an API key from app.anchord.ai/signup.

Does Anchord write to my CRMs?

No. Anchord is strictly read-only. It reads data from connected systems (Salesforce, HubSpot, Stripe) to build identity graphs, but never writes back. guard_write returns a decision — the caller performs any actual write.

What systems does Anchord work with?

Anchord has OAuth integrations for Salesforce, HubSpot, and Stripe. You can also push records from any system via the ingest_record tool or the REST API.

What happens when there's ambiguity?

When resolve_* returns needs_review, it means multiple candidate AnchorIDs matched with similar confidence. The agent should stop, surface the candidates to a human, and direct them to the Anchord Review Queue. Once resolved, subsequent calls return resolved.

What are the rate limits?

120 requests/minute per tenant. Batch endpoints accept up to 200 items (resolve, guard) or 100 records (ingest). Plan-level monthly and daily quotas apply. See docs/auth.md.

Links

• Anchord website

• Sign up

• API docs (Swagger)

• Agent quickstart

• Tool reference

• Authentication

License

MIT