GTM MCP Server

[!IMPORTANT] The hosted instance at https://mcp.gtmeditor.com will be shut down on November 30, 2026. See #139 for details. Thanks to everyone who used it!

GTM MCP Server connects MCP clients to the Google Tag Manager API. It can inspect containers, create and update workspace entities, create versions, and publish a selected version after explicit confirmation.

Use the hosted server at:

https://mcp.gtmeditor.com

The server supports browser-based Google OAuth for individual users and service-account authentication for self-hosted automation.

Project status

The project implements 101 methods from Google's 106-method GTM v2 discovery surface. The five accounts.user_permissions methods are intentionally excluded because granting and revoking GTM access are security-related and it shouldn't be manager by an LLM.

Connect an MCP client

Add the hosted URL as a remote HTTP MCP server. The client should discover the OAuth metadata, open Google sign-in, and reconnect with the issued bearer token.

Claude Code

claude mcp add --transport http gtm https://mcp.gtmeditor.com

Gemini CLI

gemini mcp add --transport http gtm https://mcp.gtmeditor.com

Gemini CLI can also use this settings.json entry:

{
  "mcpServers": {
    "gtm": {
      "httpUrl": "https://mcp.gtmeditor.com"
    }
  }
}

Cursor

Add the following to .cursor/mcp.json:

{
  "mcpServers": {
    "gtm": {
      "url": "https://mcp.gtmeditor.com"
    }
  }
}

ChatGPT, Codex, and Claude web

Add a custom or remote MCP connection in the client and use https://mcp.gtmeditor.com as the server URL. Product menus change more often than this server, so refer to the client's current MCP connection instructions if the label differs.

After connecting, try:

List my GTM accounts, containers, and workspaces. Do not make changes.

The Google account used during OAuth determines which GTM accounts the server can access.

Typical workflow

1. Call list_accounts, list_containers, and list_workspaces to discover

IDs. Do not guess IDs.

1. Inspect the workspace with list/get tools or the audit_container prompt.
2. Create or update tags, triggers, variables, templates, clients, or

transformations in the selected workspace.

1. Call get_workspace_status and resolve conflicts before versioning.
2. Call create_version to snapshot the workspace.
3. Inspect the saved version with get_version.
4. Call publish_version with confirm: true only when the selected version

is ready to go live.

1. Verify the published state with get_live_version.

Example requests:

• "Audit this workspace for duplicate tags and unused triggers."
• "Create a GA4 purchase event tag, but do not publish it."
• "Show the difference between the latest saved version and the live version."
• "Generate a Markdown tracking plan from this workspace."
• "Import the iubenda template from the Community Template Gallery."

Tools

All GTM tools use the authenticated Google identity from the current MCP request. Inputs and outputs are structured JSON.

Utility

Accounts and containers

Workspaces

Bulk update and conflict resolution accept raw GTM Entity JSON so every entity type supported by the official API remains available.

Tags

Triggers

For update_trigger, omit filterJson, customEventFilterJson, autoEventFilterJson, or parameterJson to preserve the current field. Pass the JSON string "[]" to clear a field, or a non-empty JSON array to replace it. For click, link-click, and form-submission triggers, use filterJson because GTM drops autoEventFilter for those trigger types.

Variables

Folders and built-in variables

Use revert_workspace_entity to discard changes to a built-in variable.

Zones

Use revert_workspace_entity to discard changes to a zone.

Environments

The environments tool group is optional. Enable it with GTM_TOOL_GROUPS=all or add environments to an explicit group list.

The Live and Latest environments are managed by GTM. Creation and deletion apply to user environments; GTM permits URL and debug updates on other types.

Destinations and Google tag configurations

These tools are in the optional destinations and gtag groups.

Container combine, tag-ID move, and destination link operations do not copy or enable user permissions. Account permission management remains outside the current parity target.

Custom templates

The full JSON examples live in the gtm://best-practices/tool-input-formats resource so every tool listing does not repeat them.

Server-side containers

Use revert_workspace_entity to discard changes to a client or transformation.

Versions and publication

Workspace reverts

The tool fetches the current entity fingerprint before calling the matching official revert method. A successful result can have existsAfterRevert: false when the entity does not exist in the latest container version.

Safety model

• Delete operations, publication, and disabling built-in variables require

confirm: true.

• Update operations fetch the current resource fingerprint and use Google's

optimistic concurrency checks.

• create_version first checks that the workspace has changes and no merge

conflicts.

• Read tools distinguish latest saved state from published live state.
• Google API errors are mapped to clearer not-found, permission, conflict, and rate-limit failures. Retryable API failures use bounded backoff.
• MCP request bodies are limited to 5 MiB. OAuth and dynamic registration

endpoints have stricter rate and body limits.

Most entity edits affect a workspace and are not live until a version is published. Account and container operations act directly on those resources. Always review the target IDs and the generated version before publication.

Resources

The server exposes two concrete resources and six URI templates:

Best-practice topics include naming and organization, safe edits, GA4 and consent, server-side containers, and detailed tool input formats.

Prompts

Prompts prepare context and instructions for the model. They do not bypass tool authentication or mutation safeguards.

Authentication

Hosted OAuth

The hosted server implements MCP OAuth discovery and Google authorization. It supports PKCE, Dynamic Client Registration, Client ID Metadata Documents, authorization-server metadata, and protected-resource metadata.

The server never receives a Google password. It receives Google OAuth tokens after consent. A self-hosted operator can keep issued MCP and Google tokens across restarts by setting TOKEN_STORE_PATH; without that setting they remain in memory and disappear when the process stops.

Expired MCP access tokens can be renewed transparently while their Google refresh credentials remain valid. AUTH_AUTO_REFRESH_MAX_AGE limits how long one bearer can be silently extended; its default is seven days.

Service-account mode

Service-account mode gives every holder of one server API key the GTM access granted to the configured Google service account.

1. Create a Google service account.
2. Add its email address to the required GTM account with only the permissions

it needs.

1. Set a strong SERVICE_ACCOUNT_API_KEY on this server.
2. Set GOOGLE_SERVICE_ACCOUNT_KEY_JSON to the key JSON, or use Application

Default Credentials on Google Cloud.

1. Configure the MCP client to send

Authorization: Bearer <SERVICE_ACCOUNT_API_KEY>.

OAuth and service-account mode can run together. Requests with the configured API key use the service account; OAuth users keep their own Google identity and GTM permissions.

Self-hosting

Requirements

• Go 1.26 or Docker
• A Google Cloud project with the Tag Manager API enabled
• A Google OAuth web client for user OAuth, or a Google service account for S2S
• HTTPS and a stable public URL for remote OAuth deployments

Google OAuth setup

Create an OAuth 2.0 Web application in Google Cloud and add this authorized redirect URI:

https://your-host.example/oauth/callback

The scheme and host must match BASE_URL exactly. For local development, use:

http://localhost:8080/oauth/callback

Run from source

git clone https://github.com/paolobietolini/gtm-mcp-server.git
cd gtm-mcp-server

cat > .env <<'EOF'
BASE_URL=http://localhost:8080
GOOGLE_CLIENT_ID=your-client-id.apps.googleusercontent.com
GOOGLE_CLIENT_SECRET=your-client-secret
TOKEN_STORE_PATH=./data/tokens.json
EOF

go run .

The repository reads .env and then .env.local; .env.local overrides .env. Both are ignored by Git.

Run with Docker

docker build -t gtm-mcp-server .

docker run --rm \
  --name gtm-mcp-server \
  -p 8080:8080 \
  --env-file .env \
  -v gtm-mcp-tokens:/data \
  gtm-mcp-server

When using the named volume, set TOKEN_STORE_PATH=/data/tokens.json in .env.

Configuration

If OAuth and service-account credentials are both absent, the server starts in open mode. Use open mode only for isolated local development.

TRUST_PROXY=true is appropriate only when a trusted reverse proxy overwrites or appends X-Forwarded-For. The implementation uses the rightmost value. With multiple proxy hops, configure and test the trust boundary before relying on per-client rate limits.

ALLOWED_HOSTS is a comma-separated allowlist used when the same server is reached through trusted internal Docker hostnames. Do not add arbitrary public hosts.

GTM_TOOL_GROUPS controls schema size for clients that need only part of the API. Available groups are accounts, workspaces, tags, triggers, variables, folders, builtins, zones, templates, server, guidance, environments, destinations, gtag, container-admin, folder-admin, and workspace-admin, version-admin, and reverts. Leaving it unset preserves the established 64-tool surface. New parity groups are opt-in. Use all to include every current and future parity family, or select a subset:

GTM_TOOL_GROUPS=accounts,workspaces,tags,triggers,variables

The two connection utility tools remain available regardless of this setting.

Optional setup dashboard

Set DASHBOARD_ENABLED=true and restart the server to expose /dashboard/. The page shows the canonical MCP endpoint, server version, authentication setup status, and connection instructions. It is disabled by default.

The dashboard is public when enabled. It deliberately excludes tokens, sessions, account names, container data, and other GTM content. It reports server configuration only and does not verify the visitor's Google session.

Releases and deployment

server.json is the source of truth for the runtime version. It is embedded in the Go binary and returned by /health; there is no second version constant in Go code.

To publish a release:

1. Update server.json.
2. Commit the release changes.
3. Push a matching tag such as v1.11.0.

The release workflow rejects a tag that does not match server.json, creates cross-platform archives with GoReleaser, and then deploys the tagged source to the production VPS. The deployment uses the GitHub environment auto-deployment and expects these secrets:

• SSH_PRIVATE_KEY
• VPS_KNOWN_HOSTS
• VPS_HOST
• VPS_USER

The workflow preserves server-only .env, docker-compose.yml, and token data, keeps a rollback image, rebuilds the service, and waits until /healthreports the expected version.

Recent commits on main can be newer than the latest tagged release. Check the release page when you need a reproducible published artifact.

Development

go test ./... -count=1
go vet ./...
staticcheck ./...
go run ./cmd/tool-schema-report
go run ./cmd/tool-schema-report -groups tags,zones

The schema report prints the GTM tool count, serialized tools/list size, token estimate, and largest definitions. The default GTM tool surface serializes to 78,734 bytes for 64 tools. A regression test enforces an 80,000-byte ceiling so new parity work does not silently consume unlimited model context. The all group exposes 94 GTM tools and serializes to 115,748 bytes.

Pull requests run govulncheck, gosec, Gitleaks, Trivy, staticcheck, and CodeQL. Request-level GTM tests use local fake Google endpoints. Mutating live tests must use a disposable container and clean up their entities.

See ARCHITECTURE.md for package boundaries, request flow, authentication internals, token persistence, and security invariants.

For a deep dive into the Google Tag Manager MCP server, read the Deep Wiki.

Additional context

• llms.txt is served at

https://mcp.gtmeditor.com/llms.txt.

• skills/gtm-mcp/gtm-mcp.md contains a Claude

Code skill with workflow guidance.

• examples/gtm_agent.py demonstrates a programmatic

MCP client.

• Google Tag Manager API v2 reference
• Model Context Protocol

License

BSD 3-Clause. See LICENSE.

Maintained by Paolo Bietolini.