Troubleshooting the GitLab MCP server

When working with the GitLab MCP server, you might encounter the following issues.

Error: 403 Forbidden

You might get this error when you start the GitLab MCP server. You might also get this error when POST /api/v4/mcp or GET /api/v4/mcp returns 403 Forbidden after the OAuth flow completes.

In GitLab 19.4 and earlier, the same issue returns 404 Not Found instead. Both status codes share the same causes.

To resolve this issue, make sure you meet the prerequisites for the GitLab MCP server.

To find the cause, read the message in the response body. Administrators can also check the denial_reason field in the mcp.log file:

• MCP server disabled for this instance (denial_reason: instance_setting_disabled): On GitLab Self-Managed, an administrator turned off the MCP server for the instance.
• MCP server not enabled for any of your groups (denial_reason: no_enabled_namespace): On GitLab.com, no top-level group you belong to has the MCP server turned on.

Error: Server's protocol version is not supported: 2025-06-18

In GitLab 18.6 and earlier, you might get this error when the MCP client library does not support the GitLab MCP server protocol specification.

To resolve this issue, ask the AI tool provider to update their client implementation.

Error: 429 Too Many Requests

Your MCP client might stop working and report a generic HTTP or connection error, and every request fails until the rate limit period resets. This happens when a rate limit is exceeded: either the MCP server rate limit or the authenticated API rate limit, because MCP server requests count against both. GitLab refuses the request before it reaches the MCP server, so no tool runs. Unlike a rate_limited tool result, this is not a 200 OK that carries an error result.

The response is HTTP 429 with content type text/plain and, by default, the body Retry later. It is not a JSON-RPC message, so an MCP client cannot read it as an MCP response.

To confirm the cause, check the response headers:

• RateLimit-Name identifies which limit was reached. On GitLab Self-Managed, the MCP server limit is throttle_authenticated_mcp. Any other value means a different limit, such as the general authenticated API limit.
• Retry-After is the number of seconds to wait before you retry. It is the time remaining in the current period, not the full period. Clients should honor it rather than retry immediately.

The response also includes the other rate limit headers.

Before the limit is reached, successful 200 OK responses already include RateLimit-Remaining, so a client or proxy that reads headers can see the limit approaching.

Administrators can also check the logs:

• A refused request appears in the Workhorse log with status 429 and uri /api/v4/mcp. The entry has no user and no limit name.
• A refused request does not appear in api_json.log or mcp.log, because it never reaches the API. It also does not appear in auth.log, unlike some other GitLab rate limits.
• Requests that were counted but allowed appear in api_json.log with a rate_limit_state field containing rack_request_mcp:authenticated_mcp:allow. Use this to identify which users consume the limit.

To resolve this issue, wait for the number of seconds in Retry-After, then retry.

• On GitLab.com, per-plan limits apply to MCP requests. For the published values, see rate limits.
• On GitLab Self-Managed, the limit is disabled by default. If it is enabled, an administrator can raise or disable it.

Error: rate_limited tool result

If you got an HTTP 429 Too Many Requests response instead, see 429 Too Many Requests.

You might get a tool result with isError: true, even though the MCP server itself returned 200 OK. This happens when a tool call hits a rate limit on the underlying GitLab API endpoint it calls, for example the search rate limit when a search tool runs.

The MCP server does not return 429 Too Many Requests for this case. Instead, it follows the Model Context Protocol specification, which classifies API failures as tool execution errors reported in the result so a client or language model can read them and self-correct.

The content field contains a human-readable message, for example:

Rate limited by search_rate_limit. Retry after 60 seconds.

The structuredContent field contains machine-readable retry detail:

{
  "error": {
    "type": "rate_limited",
    "retry_after_seconds": 60,
    "limit": "search_rate_limit",
    "message": "This endpoint has been requested too many times. Try again later."
  }
}

To resolve this issue, wait and retry the tool call. When you build an MCP client, check error.type == "rate_limited" rather than match on the message text, because the wording can change. Use limit to identify which rate limit was hit, as a single tool call can consume more than one. Treat retry_after_seconds as a safe upper bound rather than an exact wait time, because it is the full rate limit period rather than the time remaining in the current window.

Error: 403 Forbidden tool result from start_duo_session

You might get an error that states 403 Forbidden when you call the start_duo_session tool. The MCP server returns 200 OK, and the tool result has isError: true with the following message in the content field:

403 Forbidden

This issue occurs when you cannot run the flow in the project. start_duo_session always starts the flow as a session that runs in a CI job, and you must have permission to run that flow in the project.

To resolve this issue, do one of the following:

• If you do not have access to GitLab Duo Agent Platform, meet the prerequisites for GitLab Duo Agent Platform.
• If you have the Reporter role or lower in the project, ask a project Owner or Maintainer for at least the Developer role.
• If foundational flows are turned off, ask a project Maintainer or Owner, or a top-level group Owner, to turn on foundational flows.
• If the flow is in beta, ask a top-level group Owner to turn on beta and experimental features.

If you do not use GitLab Duo Agent Platform, you can hide its tools from your MCP client. If your client supports custom headers, send the X-Gitlab-Enabled-Mcp-Server-Toolsets header with a list that does not include duo_agent_platform. For more information, see select tool groups (toolsets).

Troubleshoot the GitLab MCP Server in Cursor

1. In Cursor, to open the Output view, do one of the following:
   • Go to View > Output.
   • In macOS, press Command+Shift+U.
   • In Windows or Linux, press Control+Shift+U.
2. In the Output view, select MCP:SERVERNAME. The name depends on the MCP configuration value. The example with GitLab results in MCP: user-GitLab.
3. When reporting bugs, copy the output into the issue template logs section.

Troubleshoot the GitLab MCP Server on the CLI with mcp-remote

1. Install Node.js version 20 or later.

2. To test the exact same command as the IDEs and desktop clients:

   1. Extract the MCP configuration.
   2. Assemble the npx command string into one line.
   3. Run the command string.
   shell

   rm -rf ~/.mcp-auth/mcp-remote*
   
   npx -y mcp-remote@latest https://gitlab.example.com/api/v4/mcp --static-oauth-client-metadata '{"scope": "mcp"}'

3. Add the --debug parameter to log more verbose output:

   shell

   rm -rf ~/.mcp-auth/mcp-remote*
   
   npx -y mcp-remote@latest https://gitlab.example.com/api/v4/mcp --static-oauth-client-metadata '{"scope": "mcp"}' --debug

4. Optional. Run the mcp-remote-client executable directly.

   shell

   rm -rf ~/.mcp-auth/mcp-remote*
   
   npx -p mcp-remote@latest mcp-remote-client https://gitlab.example.com/api/v4/mcp --static-oauth-client-metadata '{"scope": "mcp"}'

5. Optional. If you encounter version-specific bugs, pin the version of the mcp-remote module to a specific version. For example, use mcp-remote@0.1.26 to pin the version to 0.1.26.

   For security reasons, you should not pin versions if possible.

Troubleshoot GitLab MCP Server with Claude Desktop

Verify the installed Node.js versions. Claude Desktop requires Node.js version 20 or later.

for n in $(which -a node); do echo "$n" && $n -v; done

Delete MCP authentication caches

The MCP authentication is heavily cached locally. While troubleshooting, you might encounter false positives. To prevent these, delete the cache directory during troubleshooting:

rm -rf ~/.mcp-auth/mcp-remote*

Debugging and development tools

MCP Inspector is an interactive developer tool for testing and debugging MCP servers. To run this tool, use the command line and access the web interface to inspect the GitLab MCP Server.

npx -y @modelcontextprotocol/inspector npx