mcp-gopls – MCP server for Go (gopls)

A Model Context Protocol (MCP) server that lets AI assistants use Go’s LSP (gopls) for navigation, diagnostics, testing, coverage, and more.

TL;DR: If you use Claude / Cursor / Copilot with Go, mcp-gopls gives the AI full LSP powers: go-to-definition, references, hover, completion, go test, coverage, go mod tidy, govulncheck, etc.

Overview

This MCP server helps AI assistants to:

• Use LSP to analyze Go workspaces

• Navigate to definitions, references, and workspace symbols

• Format, rename, and inspect code actions without leaving MCP

• Run Go tests, coverage, go mod tidy, govulncheck, and module graph commands with structured results

• Read workspace resources (overview + go.mod) and consume curated prompts

Status: Actively developed – used in real projects.
Tested with Go 1.25.x and gopls@latest.

Related MCP server: mcp-server-leetcode

Architecture

This project uses the mark3labs/mcp-go library to implement the Model Context Protocol. The MCP integration enables seamless communication between AI assistants and Go tools.

The server communicates with gopls, the official language server for Go, via the Language Server Protocol (LSP).

Features

• Configurable runtime: --workspace, --gopls-path, --log-level, --rpc-timeout, and --shutdown-timeout flags + env vars (MCP_GOPLS_*)

• Structured logging: Text/JSON logging with slog and optional file output

• Extended LSP surface: navigation, diagnostics, formatting, rename, code actions, hover, completion, workspace symbols

• Test & tooling helpers: coverage analysis, go test, go mod tidy, govulncheck, go mod graph

• MCP extras: resources (resource://workspace/overview, resource://workspace/go.mod) and prompts (summarize_diagnostics, refactor_plan)

• Progress streaming: long-running commands emit notifications/progress events so clients can surface status updates

Feature comparison: mcp-gopls vs built-in gopls MCP

As of gopls v0.20.0, the built-in MCP server exposes these tools: go_context, go_diagnostics, go_file_context, go_file_diagnostics, go_file_metadata, go_package_api, go_references, go_rename_symbol, go_search, go_symbol_references, go_workspace, go_vulncheck.

If you want full LSP-like editing + tooling from MCP (definition, hover, completion, format, rename, code actions, go test, coverage, go mod tidy, module graph), mcp-gopls is strictly richer.

If you mostly want read-only/introspective tools (diagnostics, symbol search, references, package API, workspace/file context, vulncheck) with no extra binary, the built-in gopls MCP is enough.

Note: The built-in gopls MCP server is still marked experimental and its tool set may change over time. This comparison is accurate as of gopls v0.20.x.

Project Structure

Installation

Note: the Go module path is github.com/hloiseaufcms/mcp-gopls, even though the GitHub repo is under hloiseau.

Quick Start

1. Install the server:

2. Verify it's on your $PATH:

3. Configure your AI client (see examples below for Cursor, Claude Desktop, or GitHub Copilot).

Detailed Client Setup

Note: All clients point to the same command:
mcp-gopls --workspace /absolute/path/to/your/go/project
The configuration format differs slightly per client, but the binary and arguments remain identical.

1. Connect from Cursor

1. Open Settings → MCP Servers → Edit JSON.

2. Add or update the mcp-gopls entry:

3. Run Developer: Reload Window so Cursor reconnects.

4. Open the Tools drawer in Cursor Chat and enable mcp-gopls.

2. Invoke the tools

Client Setup Examples

Claude Desktop (macOS, Windows, Linux)

1. Install mcp-gopls and make sure it is on your $PATH.

2. Create or edit claude_desktop_config.json.

   • macOS: ~/Library/Application Support/Claude/claude_desktop_config.json

   • Linux: ~/.config/Claude/claude_desktop_config.json

   • Windows: %APPDATA%\Claude\claude_desktop_config.json

3. Add the server entry:

Restart Claude Desktop, open a chat, and ask it to connect to the mcp-gopls tool (Claude will show a “Tools” tab once the server is detected). Typical prompts include “list diagnostics for cmd/api/server.go” or “rename userService to accountService.”

Cursor IDE

In Cursor open Settings → MCP Servers → Edit JSON (this writes to ~/.cursor/config.json or the project-local override). Add:

Reload Cursor (or run the Developer: Reload Window command) and the server will appear inside the “Tools” drawer. You can now ask Cursor Chat things like “run go test ./pkg/server with coverage” or “show hover info for pkg/tools/tests.go:42.”

GitHub Copilot (Agent Mode)

GitHub Copilot’s Agent Mode can talk to local MCP servers across VS Code, JetBrains IDEs, Eclipse, and Xcode (docs). To wire mcp-gopls in VS Code:

1. Update GitHub Copilot (requires VS Code 1.99+), opt into Agent Mode.

2. Create .vscode/mcp.json in your workspace (or edit the global file shown in the Copilot “Edit config” dialog).

3. Add:

4. Reload Agent Mode (toggle off/on) so Copilot discovers the new tool; the chat “Tools” picker will now expose every MCP action (run_go_test, run_govulncheck, etc.). JetBrains and other IDEs share the same JSON schema via their Copilot settings panel.

MCP Inspector / CLI testing

For quick smoke tests or demos you can use mark3labs/mcp-inspector:

The inspector lets you call each tool/resource/prompt manually, which is handy for debugging server configuration before wiring it into an AI assistant.

MCP Tools

Progress Notifications

Long-running tools emit structured notifications/progress events so IDEs can show rich status indicators:

• Streaming progress (run_go_test, analyze_coverage, run_govulncheck, run_go_mod_tidy) forwards incremental log lines and percentage updates. Cursor displays these as a live log.

• Start/complete events only (go_to_definition, find_references, rename_symbol, etc.) fire a quick “started” event so the UI can show a spinner, followed by a completion payload with the final result.

• Each progress token is now namespaced (e.g., run_go_test/<rand>) to avoid “unknown token” errors when multiple tools run concurrently.

When integrating new tools, opt into streaming mode only if the underlying LSP/golang command produces meaningful interim output; otherwise stick to the lightweight start/complete flow to minimize noise.

Prompt Instructions

Both prompts are accessible from any MCP-aware client via the “Prompts” catalog.

summarize_diagnostics

• When to use: After check_diagnostics or run_go_test to turn raw diagnostics into actionable steps.

• Arguments: None. The server automatically reads the last diagnostics payload cached by the tools layer.

• Typical workflow: check_diagnostics → copy the returned array into the prompt input field (Cursor’s UI pastes it automatically when you select “Use last result”).

refactor_plan

• When to use: You already have a diagnostics JSON array and want a concise change checklist.

• Arguments: Requires a diagnostics object containing the raw Go diagnostics (the same payload returned by check_diagnostics).

• Example invocation payload:

The prompt responds with a numbered set of refactor steps plus suggested validation commands (go test, analyze_coverage, etc.).

Configuration

The server supports various configuration options via command-line flags and environment variables:

Command-Line Flags

Environment Variables

All flags can be set via environment variables with the MCP_GOPLS_ prefix:

Command-line flags take precedence over environment variables.

Troubleshooting

• “column is beyond end of line” – gopls could not map the provided position. Confirm the file is saved and the position uses zero-based lines/columns; run go fmt to ensure tabs vs. spaces align with gopls expectations.

• “no hover information available” – the symbol might belong to a generated file or a module outside the configured workspace. Ensure the --workspace flag points to the module root and that go list ./... succeeds.

• “workspace not initialized” – the server did not finish its initial sync. Wait for the workspace initialized log line or restart mcp-gopls after deleting stale .gopls caches.

• run_govulncheck – the tool now falls back to go run golang.org/x/vuln/cmd/govulncheck@latest, but the machine still needs outbound network access. Install the binary manually if the fallback is blocked.

Usage Example

Using the server with AI assistants that support MCP:

Development

Table-driven tests live under pkg/tools and CI runs via .github/workflows/ci.yml.

Documentation

• docs/usage.md – quickstart and tool catalog walkthrough

• Workspace resources expose resource://workspace/overview and resource://workspace/go.mod

• Prompts (summarize_diagnostics, refactor_plan) help assistants produce consistent outputs

Contributing

PRs and issues are welcome!

• Check open issues or file a new one if you hit a bug or want a feature.

• Run go test ./... before opening a PR.

• For bigger changes (new tools, protocol changes), please open a design issue first so we can discuss the approach.

All contributions should maintain test coverage and adhere to Go best practices. See the Development section for setup instructions.

Prerequisites

• Go 1.25+ (tested with go1.25.4)

• gopls installed (go install golang.org/x/tools/gopls@latest)

• Optional: govulncheck (go install golang.org/x/vuln/cmd/govulncheck@latest)

• The server forces GOTOOLCHAIN=local for its nested gopls process. If you need a different toolchain, set GOTOOLCHAIN in the environment before launching mcp-gopls.

Integration with Ollama

This MCP server can be used with any tool that supports the MCP protocol. For Ollama integration:

1. Make sure Ollama is running

2. The MCP server runs independently and communicates through stdin/stdout

3. Configure your client to use the MCP server as a tool provider

License

Apache License 2.0