Fingerprint MCP Server

Features

• Run locally, or use the managed instance at https://mcp.fpjs.io/mcp (in Claude, add it from the connectors directory)

• Event tools: Retrieve and search identification and Automation Intelligence (edge) events with full smart signal data

• Management tools: Manage workspace environments, API keys and custom subdomains

• Skills tools: Serve the Fingerprint skills (integration guides, proxy setup, Smart Signals) to clients that cannot fetch them themselves

• Supports both stdio and streamable-http transports

• Optional HTTPS with TLS certificates

• OAuth2 login supported

• Public mode for multi-tenant deployments (API keys passed via JWT bearer token)

• Tool filtering to control which tools are exposed

• Configurable via environment variables or CLI flags

• Docker support for easy deployment

Related MCP server: Enterprise MCP Gateway and Tool Registry

Installation

From Source

go install github.com/fingerprintjs/fingerprint-mcp-server@latest

Build Locally

git clone https://github.com/fingerprintjs/fingerprint-mcp-server.git
cd fingerprint-mcp-server
make build

make build runs go generate (to download the OpenAPI spec and generate schema files under internal/schema/) and then builds the binary.

Configuration

The server can be configured via CLI flags or environment variables:

Telemetry

This binary ships with telemetry disabled. There is no built-in analytics backend and the default emitter is a no-op, so running the OSS binary sends nothing to anyone.

The server exposes a hook (analytics.Emitter in the analytics package) that fires on each MCP method. Embedders who want product analytics register their own emitter by importing this package as a library and calling WithAnalytics(emitter) when constructing the server. Fingerprint's hosted build (mcp.fpjs.io) does this with an Amplitude emitter, but that implementation lives in the private managed repo, not here.

If you build your own emitter, the event shape is in analytics/analytics.go. The middleware fires one mcp_method_called event per MCP method, gated on a non-empty subscription_id (so pre-auth methods and the private-mode auth path stay silent regardless). Calls that go through call_tool or call_write_tool report the proxy as tool_name and the tool it was asked to run as target_tool, also when the proxy rejects the call (then is_error is true).

Request inspection

Embedders can register a second hook (requestinspect.Inspector in the requestinspect package) that receives HTTP request metadata — method, URL, original headers, and TCP peer address — for every request to the /mcp endpoint, useful for audit logging and traffic analysis. Register it with WithRequestInspector(inspector) when constructing the server. The OSS binary configures no inspector, so nothing is collected.

The full contract (timing, fail-open behavior, read-only fields, stdio) is documented in requestinspect/requestinspect.go, the source of truth for this hook. See requestinspect/example_test.go for a complete embedder-side implementation skeleton (bounded queue, background delivery, header redaction, graceful drain).

Usage

Private mode vs. Public mode

Private mode means the server runs with its API keys pre-configured (FINGERPRINT_SERVER_API_KEY and FINGERPRINT_MANAGEMENT_API_KEY). This mode is useful when you are running a local instance intended to be used within your organization: server automatically uses those specified in the config. In this mode, auth token (MCP_AUTH_TOKEN) is enforced to protect your instance of the MCP server from unauthenticated use.

Public mode is how https://mcp.fpjs.io/mcp is run. It is meant to be used in situations when a single instance can be used by different users from different organizations, each with their own API keys. In this mode, API keys are extracted from JWT access tokens that are issued by https://dashboard.fingerprint.com or by user following the OAuth2 flow.

Stdio Transport (Default)

export FINGERPRINT_SERVER_API_KEY=your-secret-api-key
./fingerprint-mcp-server

Streamable HTTP Transport

export FINGERPRINT_SERVER_API_KEY=your-secret-api-key
./fingerprint-mcp-server --transport=streamable-http --port=8080

The MCP endpoint will be available at http://localhost:8080/mcp.

HTTPS

Provide TLS certificate and key files to enable HTTPS:

./fingerprint-mcp-server --transport=streamable-http \
  --tlscert=cert.pem --tlskey=key.pem

Tool Filtering

By default, all tools are registered based on which API keys are configured. You can restrict which tools are exposed:

# Only expose read-only tools
./fingerprint-mcp-server --readonly

# Expose a specific set of tools
./fingerprint-mcp-server --tools=get_event,search_events,list_environments

When --tools is set, it overrides --readonly.

Docker

Build the Image

docker build -t fingerprint-mcp-server .

Run with Stdio Transport

docker run -i --rm \
  -e FINGERPRINT_SERVER_API_KEY=your-secret-api-key \
  fingerprint-mcp-server

Run with Streamable HTTP Transport

docker run -d --rm \
  -e FINGERPRINT_SERVER_API_KEY=your-secret-api-key \
  -e MCP_TRANSPORT=streamable-http \
  -p 8080:8080 \
  fingerprint-mcp-server

MCP Client Configuration

Hosted endpoint (managed)

The managed instance at https://mcp.fpjs.io/mcp runs in public mode and authenticates each user via OAuth2. Nothing to run or configure.

In Claude, add it from the connectors directory rather than as a custom connector: open Customize > Connectors, choose Browse connectors, and search for Fingerprint. The same catalog serves claude.ai, Desktop, mobile and Claude Code, so this only has to be done once per account.

Adding it as a custom connector still works, but you have to name it yourself, and some characters in that name break tool discovery while the connector still reports as connected.

In other clients, point them at the URL directly:

{
  "mcpServers": {
    "fingerprint": {
      "type": "http",
      "url": "https://mcp.fpjs.io/mcp"
    }
  }
}

Self-hosted

Add to your Cursor/Claude Desktop/etc configuration file (e.g. claude_desktop_config.json):

Stdio Transport:

{
  "mcpServers": {
    "fingerprint": {
      "command": "/path/to/fingerprint-mcp-server",
      "env": {
        "FINGERPRINT_SERVER_API_KEY": "your-secret-api-key"
      }
    }
  }
}

Streamable HTTP Transport:

{
  "mcpServers": {
    "fingerprint": {
      "url": "https://url/mcp",
      "headers": {
        "Authorization": "Bearer <auth-token>"
      }
    }
  }
}

Docker with Stdio:

{
  "mcpServers": {
    "fingerprint": {
      "command": "docker",
      "args": ["run", "-i", "--rm", "-e", "FINGERPRINT_SERVER_API_KEY=your-secret-api-key", "fingerprint-mcp-server"]
    }
  }
}

Available Tools

Event tools require a Server API key. Management tools require a Management API key. get_current_time and the skills tools need no key. get_skill and get_skill_file are not in the MCP tool list: run them through call_tool. With --tools, allow-list list_skills, get_skill, get_skill_file and call_tool for the skills to work. Write tools (create/update/delete/verify/connect) are hidden when --readonly is set or excluded via --tools.

For custom subdomains, follow fingerprint-proxy-integration through the skills tools. Creation does not add DNS records: when the provider supports Domain Connect, connect_subdomain returns a link the user opens to add them with one approval (they land on the Fingerprint dashboard afterwards); otherwise the user or an authorized DNS-provider tool adds them. Configure the app only when the subdomain is active. Use get_subdomain for status checks after verifying; respect retry_after on rate-limit errors. Subdomain deletion is not exposed.

Available Resources

• fingerprint://events/{event_id} — Returns full event data for a given event ID.

• fingerprint://schemas/event — JSON Schema describing the event output structure.

• fingerprint://schemas/environment — JSON Schema for environment objects.

• fingerprint://schemas/api-key — JSON Schema for API key objects.

Skills

The skills tools read the Fingerprint skills repo from GitHub and cache each file for an hour. Pass WithSkillsRepo("", "") to disable them on a deployment with no egress.

License

MIT