GitLab MCP Server

A Model Context Protocol (MCP) server that exposes the entire GitLab API as MCP tools, resources, and prompts for AI assistants. Single static binary — zero dependencies.

Security first: Continuously monitored on SonarCloud with quality gates, coverage, and security scanning. Supports read-only mode, safe mode (dry-run preview), and self-hosted GitLab with TLS verification.

Repository mirror: GitHub is the canonical repository. A read-only mirror of the code and releases is available on GitLab.com for discoverability; please open code contributions on GitHub.

Token Footprint

Measured with go run ./cmd/gen_readme/ against the current base catalog. Totals estimate startup context visible to an MCP client: visible tool schemas plus shared resources and prompts, using the same byte/4 token heuristic as cmd/audit_tokens.

Default configuration: with TOOL_SURFACE unset or TOOL_SURFACE=dynamic, CAPABILITY_SURFACE=full, META_TOOLS unset, META_PARAM_SCHEMA=opaque, and GITLAB_ENTERPRISE unset or false, the server uses the dynamic find/execute surface. Use TOOL_SURFACE=meta only when you explicitly want domain meta-tools; use TOOL_SURFACE=individual only when your client can handle the full tool catalog.

Configuration (TOOL_SURFACE / CAPABILITY_SURFACE)	Visible tools	Reachable actions	META_PARAM_SCHEMA	Tool schema tokens	Shared tokens	Total tokens
dynamic / full (default)	2	870	n/a	2,193	18,305	20,498
dynamic / minimal	2	870	n/a	2,193	184	2,377
meta / full	34	870	opaque	66,109	18,305	84,414
meta / minimal	34	870	opaque	66,109	760	66,869
individual / full	866	866	n/a	473,192	17,729	490,921

Rows use the base Community Edition catalog (GITLAB_ENTERPRISE=false). META_PARAM_SCHEMA=opaque affects only visible meta-tool input schemas; dynamic mode gets exact action schemas from gitlab_find_action, and individual mode already exposes one schema per tool. In meta + minimal, shared tokens still include meta-schema resources so opaque meta-tools can look up exact action parameter schemas.

Highlights

• 1017 MCP tools on self-managed Enterprise/Premium, or 1022 on GitLab.com Enterprise/Premium with experimental Orbit Knowledge Graph support — broad GitLab REST API v4 + GraphQL coverage across 165 domain sub-packages: projects, branches, tags, releases, merge requests, issues, pipelines, jobs, groups, users, wikis, environments, deployments, packages, container registry, runners, feature flags, CI/CD variables, security attributes, security categories, templates, admin settings, access tokens, deploy keys, Orbit, and more
• Default dynamic toolset — exposes only gitlab_find_action and gitlab_execute_tool while keeping the same canonical GitLab action catalog. Optional domain meta-tools remain available with TOOL_SURFACE=meta: 33 base, 49 on self-managed Enterprise/Premium, or 50 on GitLab.com Enterprise/Premium
• AI model tool-use evaluation — automated schema-only and Docker-backed runs against a populated GitLab CE instance measure tool/action selection, parameter shaping, recovery from GitLab errors, and destructive-action safety across Anthropic, Google, OpenAI, and Qwen. Published summaries appear in the managed evaluation block below; see AI Model Evaluation Results
• 11 sampling actions — LLM-assisted code review, issue analysis, pipeline failure diagnosis, security review, release notes, milestone reports, and more via gitlab_analyze meta-tool (MCP sampling capability)
• 4 elicitation tools — interactive creation wizards (issue, MR, release, project) with step-by-step user prompts
• 46 MCP resources — read-only data: user, groups, group members, group projects, projects, issues, pipelines, members, labels, milestones, branches, MRs, releases, tags, commits, file blobs, wiki pages, MR notes, MR discussions, meta-tool JSON Schemas, single-entity templates (issue, MR, branch, tag, release, label, milestone, commit, wiki page, deployment, environment, job, board, snippet, deploy key, feature flag, group label, group milestone), workspace roots, and 5 workflow best-practice guides
• 37 MCP prompts — AI-optimized: code review, pipeline status, risk assessment, release notes, standup, workload, user stats, team management, cross-project dashboards, analytics, milestones, Git workflow quality, audit
• 6 MCP capabilities — logging, completions, roots, progress, sampling, elicitation
• 50 tool icons — base64-encoded SVG icons (Sizes: ["any"]) on all tools, resources, and prompts for visual identification in MCP clients
• Pagination on all list endpoints with metadata (total items, pages, next/prev)
• Transports: stdio (default for desktop AI) and HTTP (Streamable HTTP for remote clients)
• Cross-platform: Windows, Linux & macOS, amd64 & arm64
• Self-hosted GitLab with self-signed TLS certificate support

Example Prompts

Once connected, just talk to your AI assistant in natural language:

"List my GitLab projects"
"Show me open merge requests in my-app"
"Create a merge request from feature-login to main"
"Review merge request !15 — is it safe to merge?"
"List open issues assigned to me"
"What's the pipeline status for project 42?"
"Why did the last pipeline fail?"
"Generate release notes from v1.0 to v2.0"

The server handles the translation from natural language to GitLab API calls. You do not need to know project IDs, API endpoints, or JSON syntax — the AI assistant figures that out for you. See Usage Examples for more scenarios.

Quick Start

1. Get the server

Download the latest binary for your platform from GitHub Releases and make it executable:

chmod +x gitlab-mcp-server-*  # Linux/macOS only

Or pull the published container image:

docker pull ghcr.io/jmrplens/gitlab-mcp-server:latest

2. Configure GitLab access

Recommended: Run the built-in setup wizard — it configures your GitLab connection and MCP client in one step:

./gitlab-mcp-server --setup

Tip: The wizard supports Web UI, Terminal UI, and plain CLI modes. On Windows, double-click the .exe to launch the wizard automatically.

Manual setup only needs a GitLab Personal Access Token with api scope:

GITLAB_TOKEN=glpat-xxxxxxxxxxxxxxxxxxxx

GITLAB_URL defaults to https://gitlab.com; add it only when you connect to a self-managed GitLab instance.

GITLAB_URL=https://gitlab.example.com

3. Connect your MCP client

Most desktop clients use stdio: the client starts one local MCP server process and talks to it over stdin/stdout. Choose one of these runtime patterns.

Native binary (stdio)

VS Code and Cursor-style MCP configuration:

Add to .vscode/mcp.json in your workspace:

{
  "servers": {
    "gitlab": {
      "type": "stdio",
      "command": "/path/to/gitlab-mcp-server",
      "env": {
        "GITLAB_TOKEN": "glpat-xxxxxxxxxxxxxxxxxxxx"
      }
    }
  }
}

Claude Desktop uses the same server command under mcpServers:

{
  "mcpServers": {
    "gitlab": {
      "command": "/path/to/gitlab-mcp-server",
      "env": {
        "GITLAB_TOKEN": "glpat-xxxxxxxxxxxxxxxxxxxx"
      }
    }
  }
}

For client-specific paths, secure token prompts, HTTP OAuth, and extra IDEs, see IDE Configuration.

Docker launched by an IDE (stdio)

If an IDE starts Docker as the MCP server process, keep docker run -i and pass --http=false after the image name. Do not publish port 8080 in this mode.

{
  "servers": {
    "gitlab": {
      "type": "stdio",
      "command": "docker",
      "args": [
        "run",
        "-i",
        "--rm",
        "-e",
        "GITLAB_TOKEN",
        "-e",
        "GITLAB_URL",
        "-e",
        "GITLAB_SKIP_TLS_VERIFY",
        "ghcr.io/jmrplens/gitlab-mcp-server:latest",
        "--http=false"
      ],
      "env": {
        "GITLAB_TOKEN": "glpat-xxxxxxxxxxxxxxxxxxxx",
        "GITLAB_URL": "https://gitlab.com",
        "GITLAB_SKIP_TLS_VERIFY": "false"
      }
    }
  }
}

Docker or binary as an HTTP MCP server

Use HTTP mode for shared, remote, or multi-user deployments. The Docker image
starts in HTTP mode by default, but the flags are shown explicitly here for
clarity. These examples publish the container port on host loopback only;
--http-addr=0.0.0.0:8080 binds inside the container.

# Fixed GitLab instance for all clients
docker run -d --name gitlab-mcp-server -p 127.0.0.1:8080:8080 \
  ghcr.io/jmrplens/gitlab-mcp-server:latest \
  --http \
  --http-addr=0.0.0.0:8080 \
  --gitlab-url=https://gitlab.com

# Multi-instance mode: clients send GITLAB-URL per request
docker run -d --name gitlab-mcp-server -p 127.0.0.1:8080:8080 \
  ghcr.io/jmrplens/gitlab-mcp-server:latest \
  --http \
  --http-addr=0.0.0.0:8080

HTTP clients authenticate each request with PRIVATE-TOKEN or Authorization: Bearer:

{
  "servers": {
    "gitlab": {
      "type": "http",
      "url": "http://localhost:8080/mcp",
      "headers": {
        "PRIVATE-TOKEN": "glpat-xxxxxxxxxxxxxxxxxxxx"
      }
    }
  }
}

In multi-instance mode, clients must also send GITLAB-URL. See HTTP Server Mode for OAuth, reverse proxy, rate limit, and server-pool details.

4. Verify

Open your AI client and try:

"List my GitLab projects"

See the Getting Started guide for detailed setup instructions.

Tool Modes

Three registration modes, controlled by TOOL_SURFACE:

Mode	Tools	Description
Dynamic Toolset (default)	2 visible tools	Low-token find/execute surface over the canonical action catalog.
Meta-Tools	33 base domain meta-tools plus the gitlab_server helper	Domain-grouped dispatchers with action parameter. Enable with TOOL_SURFACE=meta; see the full 33/49/50 catalog in Meta-Tools Reference.
Individual	866 CE / 1017 self-managed enterprise / 1022 GitLab.com Enterprise	Every GitLab operation as a separate MCP tool.

For dynamic experiments where resources and prompts dominate initial context, set CAPABILITY_SURFACE=minimal (stdio) or --capability-surface=minimal (HTTP). Dynamic minimal keeps only gitlab://workspace/roots; meta minimal keeps workspace roots plus meta-schema resources so opaque meta-tools can still read exact action schemas. The default remains full.

Dynamic mode is now the default low-token find/execute surface; see Dynamic Toolset for the field-aware ranking model, fuzzy fallback, response shapes, workflow diagrams, and migration guidance. Set TOOL_SURFACE=meta to use the consolidated domain meta-tool catalog.

The detailed meta-tool catalog now lives in Meta-Tools Reference, including action counts, Enterprise/Premium markers, and examples.

Compatibility

MCP Capability	Support
Tools	Up to 1022 individual / 33–50 meta
Resources	46 (static + templates)
Prompts	37 templates
Completions	Project, user, group, branch, tag
Logging	Structured (text/JSON) + MCP notifications
Progress	Tool execution progress reporting
Sampling	11 LLM-powered analysis actions via gitlab_analyze
Elicitation	4 interactive creation wizards
Roots	Workspace root tracking

Tested with: VS Code + GitHub Copilot, Claude Desktop, Claude Code, Cursor, Windsurf, JetBrains IDEs, Zed, Kiro, Cline, Roo Code.

See the full Compatibility Matrix for detailed client support.

AI Model Tool-Use Evaluation

The project includes an automated evaluator for model-facing MCP quality. It can
run schema-only checks against the tool catalog or execute validated model tool
calls through MCP against a Docker GitLab CE instance populated with fixtures.
The evaluator measures whether each model chooses the correct meta-tool and
action, sends valid parameters, recovers from actionable GitLab errors, and
respects destructive-action safeguards.

Current published result: 2026-05-18 Meta and Dynamic Docker full run.

Provider	Model	Compatibility	Tool accuracy	Recovery	Docker live status
Anthropic	claude-haiku-4-5-20251001	Review	100.0%	No repairs	98.5% final across 264 ops
Google	gemini-3.1-flash-lite-preview	Review	91.2%	76.9% (10/13)	97.8% final across 264 ops
OpenAI	gpt-5.4-nano	OK	100.0%	100.0% (2/2)	100.0% final across 264 ops
Qwen	qwen3.6-flash	Review	97.0%	100.0% (2/2)	97.0% final across 264 ops

The published model-evaluation set covers 544 task attempts and 1056 expected MCP operations. Across the selected reports, models emitted 1043 tool calls over 1064 model requests, with 98.3% aggregate final success. See AI Model Evaluation Results for the detailed current matrix.

Current published result: 2026-05-18 Meta and Dynamic Docker full run.

Provider	Model	Compatibility	Tool accuracy	Recovery	Docker live status
Anthropic	claude-haiku-4-5-20251001	Review	100.0%	100.0% (1/1)	100.0% final across 256 ops
Google	gemini-3.1-flash-lite-preview	Review	99.3%	100.0% (1/1)	100.0% final across 256 ops
OpenAI	gpt-5.4-nano	OK	100.0%	No repairs	100.0% final across 256 ops
Qwen	qwen3.6-flash	OK	100.0%	No repairs	100.0% final across 256 ops

The published model-evaluation set covers 536 task attempts and 1024 expected MCP operations. Across the selected reports, models emitted 1072 tool calls over 1073 model requests, with 100.0% aggregate final success. See AI Model Evaluation Results for the detailed current matrix.

Documentation

Full documentation is available at jmrplens.github.io/gitlab-mcp-server.

Document	Description
Getting Started	Download, setup wizard, per-client configuration
IDE Configuration	Per-client stdio, HTTP legacy, and HTTP OAuth examples
Configuration	Environment variables, transport modes, TLS
HTTP Server Mode	Shared HTTP deployments, authentication, server pool isolation
Tools Reference	All individual tools with input/output schemas, including GitLab.com-only Orbit
Meta-Tools	33/49/50 domain meta-tools with action dispatching
Dynamic Toolset	2-tool low-token mode with canonical action catalog, safety model, and examples
Resources	All 46 resources with URI templates
Prompts	All 37 prompts with arguments and output format
Auto-Update	Self-update mechanism, modes, and release format
Testing	Unit, E2E, schema model evaluation, Docker model evaluation, and curated model results
Security	Security model, token scopes, input validation
Architecture	System architecture, component design, data flow
Development Guide	Building, testing, CI/CD, contributing

Tech Stack

Component	Technology
Language	Go 1.26+
MCP SDK	github.com/modelcontextprotocol/go-sdk v1.6.0
GitLab Client	gitlab.com/gitlab-org/api/client-go/v2 v2.27.0
Transport	stdio (default), HTTP (Streamable HTTP)

Building from Source

git clone https://github.com/jmrplens/gitlab-mcp-server.git
cd gitlab-mcp-server
make build

See the Development Guide for cross-compilation and contributing guidelines.

Container Image

The published image is ghcr.io/jmrplens/gitlab-mcp-server:latest. Runtime examples live in Quick Start next to MCP client configuration, and Docker Compose/source-build details live in the Development Guide.

FAQ

Does it work with self-hosted GitLab?

Yes. Set GITLAB_URL to your instance URL. When GITLAB_URL is omitted, stdio mode uses https://gitlab.com. Self-signed TLS certificates are supported via GITLAB_SKIP_TLS_VERIFY=true.

Is my data safe?

The server runs locally on your machine (stdio mode) or on your own infrastructure (HTTP mode). No data is sent to third parties — all API calls go directly to your GitLab instance. See SECURITY.md for details.

Can I use it in read-only mode?

Yes. Set GITLAB_READ_ONLY=true to disable all mutating tools (create, update, delete). Only read operations will be available.

Alternatively, set GITLAB_SAFE_MODE=true for a dry-run mode: mutating tools remain visible but return a structured JSON preview instead of executing. Useful for auditing, training, or reviewing what an AI assistant would do.

What GitLab editions are supported?

Both Community Edition (CE) and Enterprise Edition (EE). Set GITLAB_ENTERPRISE=true in stdio mode to enable additional tools for Premium/Ultimate features (DORA metrics, vulnerabilities, compliance, etc.). In HTTP mode, --enterprise can force the Enterprise/Premium catalog, otherwise CE/EE is detected per token+URL pool entry when GitLab reports edition.

How does it handle rate limiting?

The server includes retry logic with backoff for GitLab API rate limits. Errors are classified as transient (retryable) or permanent, with actionable hints in error messages.

Which AI clients are supported?

Any MCP-compatible client: VS Code + GitHub Copilot, Claude Desktop, Cursor, Claude Code, Windsurf, JetBrains IDEs, Zed, Kiro, and others. The built-in setup wizard can auto-configure most clients.

Contributing

See CONTRIBUTING.md for development guidelines, branch naming, commit conventions, and pull request process.

Security

See SECURITY.md for the security policy and vulnerability reporting.

Code of Conduct

See CODE_OF_CONDUCT.md. This project follows the Contributor Covenant v2.1.

Unnecessary Statistics

Numbers nobody asked for, but here they are anyway.

File counts

Category	Files	Lines
Source (.go, non-test)	849	142,704
Unit tests (_test.go)	444	243,703
End-to-end tests	111	24,416
Total	1,404	410,823

Functions

Category	Count
Source functions	5,634
— exported (public)	2,392
— unexported (private)	3,242
Unit test functions (TestXxx)	9,955
Subtests (t.Run(...))	2,191
End-to-end test functions	252

Ratios worth noting

Observation	Value
Test lines vs source lines	1.71× more tests than code
Average source file length	~168 lines
Average test file length	~548 lines
Comment lines in source	11,765 (~8.2% of source)
Test functions per source function	1.8×

Code patterns

Pattern	Count
if err != nil checks	5,929
defer statements	757
struct types defined	2,210
//nolint suppressions	48
TODO / FIXME / HACK comments	1

Project

Metric	Value
Go packages	207
Direct dependencies (go.mod)	11
Indirect dependencies	49
Git commits	149
Unique contributors	2

Hall of fame

Record	File
Longest source file	cmd/eval_mcp_surfaces/main.go — 8,841 lines
Longest test file	internal/tools/projects/projects_test.go — 6,903 lines

Because why not

Fact	Value
Source code printed at 55 lines/page	~2,594 pages of A4
Source lines mentioning "gitlab"	8,327 (impossible to avoid)
Longest function name in source	assertDynamicCompatibilityPolicyOwnedByActionCompat (51 chars)
Longest test function name	TestRequiredMissingAndUnknownParamNames_SchemaValidation_ReturnsSortedMissingAndUnknown (87 chars)