High-performance code intelligence MCP server. Indexes codebases into a persistent knowledge graph — average repo in milliseconds. 158 languages, sub-ms queries, 99% fewer tokens. Single static binary, zero dependencies.
MCP server for OpenCode in Claude, Cursor, and Windsurf
opencode-mcp sits between an MCP client and OpenCode’s headless API. Your client sends tool calls over stdio, and this server turns them into OpenCode requests, auto-starting an OpenCode server when needed. It also supports multiple project roots through a `directory` parameter on each tool.
Builders who want their MCP client to delegate coding work to OpenCode across one or many projects.
You can hand off features, debugging, refactors, and test runs to OpenCode without leaving your agent or editor.
What it does
Workflow tools
High-level tools like `opencode_ask`, `opencode_run`, `opencode_fire`, `opencode_check`, and `opencode_review_changes` cover common multi-step tasks.
Session management
Tools for creating, listing, polling, forking, sharing, aborting, and reverting sessions.
File and search access
Tools for reading files, finding files or symbols, and searching text or regex across a project.
Resources
Readable endpoints like `opencode://project/current`, `opencode://config`, `opencode://sessions`, and `opencode://health`.
Guided prompts
Selectable prompt templates for code review, debugging, project setup, implementation, best practices, and session summaries.
Multi-project support
Every tool can target a specific project with the `directory` parameter, so one MCP server can work across many roots.
How to get it
- 1Claude Code
claude mcp add opencode -- npx -y opencode-mcp
README
opencode-mcp
Give any MCP client the power of OpenCode.
opencode-mcp is an MCP server that bridges your AI tools (Claude, Cursor, Windsurf, VS Code, etc.) to OpenCode's headless API. It lets your AI delegate real coding work — building features, debugging, refactoring, running tests — to OpenCode sessions that autonomously read, write, and execute code in your project.
80 tools | 10 resources | 6 prompts | Multi-project | Auto-start
Why Use This?
- Delegate coding tasks — Tell Claude "build me a REST API" and it delegates to OpenCode, which creates files, installs packages, writes tests, and reports back.
- Parallel work — Fire off multiple tasks to OpenCode while your primary AI keeps working on something else.
- Any MCP client — Works with Claude Desktop, Claude Code, Cursor, Windsurf, VS Code Copilot, Cline, Continue, Zed, Amazon Q, and any other MCP-compatible tool.
- Zero setup — The server auto-starts the OpenCode HTTP server in-process via the official
@opencode-ai/sdkif one isn't already running. No manual steps.
Quick Start
Prerequisite: OpenCode must be installed.
curl -fsSL https://opencode.ai/install | bashornpm i -g opencode-aiorbrew install sst/tap/opencode
Claude Code:
claude mcp add opencode -- npx -y opencode-mcp
Claude Desktop / Cursor / Windsurf / Cline / Continue (add to your MCP config):
{
"mcpServers": {
"opencode": {
"command": "npx",
"args": ["-y", "opencode-mcp"]
}
}
}
That's it. Restart your client and OpenCode's tools will be available.
See Configuration for all client configs (VS Code Copilot, Zed, Amazon Q, etc.) and environment variables.
How It Works
MCP Client <--stdio--> opencode-mcp <--HTTP--> OpenCode Server
(Claude, Cursor, etc.) (this package) (in-process via @opencode-ai/sdk,
or external opencode serve)
Your MCP client calls tools over stdio. This server translates them into HTTP requests to the OpenCode headless API. If no OpenCode server is reachable at OPENCODE_BASE_URL, one is started in-process via the official @opencode-ai/sdk. The directory parameter on every tool routes that request to a specific project via the x-opencode-directory header, so a single MCP instance can fan out across many project roots.
Key Tools
The 80 tools are organized into tiers. Start with the workflow tools — they handle the common patterns in a single call.
Workflow Tools (13) — Start Here
| Tool | What it does |
|---|---|
opencode_setup | Check server health, providers, and project status. Use first. |
opencode_ask | Create session + send prompt + get answer. One call. |
opencode_reply | Follow-up message in an existing session |
opencode_run | Send a task and wait for completion (session + async send + polling) |
opencode_fire | Fire-and-forget: dispatch a task, return immediately |
opencode_check | Compact progress report for a running session (status, todos, files changed) |
opencode_conversation | Get formatted conversation history |
opencode_sessions_overview | Quick overview of all sessions |
opencode_context | Project + VCS + config + agents in one call |
opencode_review_changes | Formatted diff summary for a session |
opencode_wait | Poll an async session until it finishes |
opencode_provider_test | Quick-test whether a provider is working |
opencode_status | Health + providers + sessions + VCS dashboard |
Recommended Patterns
Quick question:
opencode_ask({ prompt: "Explain the auth flow in this project" })
Build something and wait:
opencode_run({ prompt: "Add input validation to POST /api/users", maxDurationSeconds: 300 })
Parallel background tasks:
opencode_fire({ prompt: "Refactor the auth module to use JWT" })
→ returns sessionId immediately
opencode_check({ sessionId: "..." })
→ check progress anytime
All Tool Categories
| Category | Count | Description |
|---|---|---|
| Workflow | 13 | High-level composite operations |
| Session | 20 | Create, list, fork, share, abort, revert, permissions |
| Message | 6 | Send prompts, execute commands, run shell |
| File & Search | 6 | Search text/regex, find files/symbols, read files |
| System | 13 | Health, VCS, LSP, MCP servers, agents, logging |
| TUI Control | 9 | Remote-control the OpenCode terminal UI |
| Provider & Auth | 6 | List providers/models, set API keys, OAuth |
| Config | 3 | Get/update configuration |
| Project | 3 | List, inspect, and initialize projects |
| Events | 1 | Poll real-time SSE events |
Resources (10)
Browseable data endpoints — your client can read these without tool calls:
| URI | Description |
|---|---|
opencode://project/current | Current active project |
opencode://config | Current configuration |
opencode://providers | Providers with models |
opencode://agents | Available agents |
opencode://commands | Available commands |
opencode://health | Server health and version |
opencode://vcs | Version control info |
opencode://sessions | All sessions |
opencode://mcp-servers | MCP server status |
opencode://file-status | VCS file status |
Prompts (6)
Guided workflow templates your client can offer as selectable actions:
| Prompt | Description |
|---|---|
opencode-code-review | Review diffs from a session |
opencode-debug | Step-by-step debugging workflow |
opencode-project-setup | Get oriented in a new project |
opencode-implement | Have OpenCode build a feature |
opencode-best-practices | Setup, tool selection, monitoring, and pitfalls |
opencode-session-summary | Summarize what happened in a session |
Multi-Project Support
Every tool accepts an optional directory parameter to target a different project. No restarts needed.
opencode_ask({ directory: "/home/user/mobile-app", prompt: "Add navigation" })
opencode_ask({ directory: "/home/user/web-app", prompt: "Add auth" })
Use opencode_project_init to scaffold a new project directory (or open a preexisting one) before the first call, so the OpenCode server registers it as a project:
opencode_project_init({ path: "/home/user/new-project" })
// → "Successfully initialized project directory at: /home/user/new-project"
opencode_run({ directory: "/home/user/new-project", prompt: "Set up a Vite + React app" })
Environment Variables
All optional. Only needed if you've changed defaults on the OpenCode server.
| Variable | Default | Description |
|---|---|---|
OPENCODE_BASE_URL | http://127.0.0.1:4096 | OpenCode server URL |
OPENCODE_SERVER_USERNAME | opencode | HTTP basic auth username |
OPENCODE_SERVER_PASSWORD | (none) | HTTP basic auth password (enables auth when set) |
OPENCODE_AUTO_SERVE | true | Auto-start an in-process OpenCode server (via @opencode-ai/sdk) if none is reachable at OPENCODE_BASE_URL |
OPENCODE_DEFAULT_PROVIDER | (none) | Default provider ID when not specified per-tool (e.g. anthropic) |
OPENCODE_DEFAULT_MODEL | (none) | Default model ID when not specified per-tool (e.g. claude-sonnet-4-5) |
Development
git clone https://github.com/AlaeddineMessadi/opencode-mcp.git
cd opencode-mcp
npm install
npm run build
npm start # run the MCP server
npm run dev # watch mode
npm test # 328 tests
Smoke Testing
End-to-end test against a running OpenCode server:
npm run build && node scripts/mcp-smoke-test.mjs
Documentation
- Getting Started — step-by-step setup
- Configuration — env vars and all client configs
- Tools Reference — all 80 tools in detail
- Resources — 10 MCP resources
- Prompts — 6 guided workflow templates
- Examples — real workflow examples
- Architecture — system design and data flow
References
License
Files in the repo
- docs
- scripts
- src
- tests
- .gitignore
- CHANGELOG.md
- CODE_OF_CONDUCT.md
- CONTRIBUTING.md
- LICENSE
- package-lock.json
- package.json
- README.md
- tsconfig.json
- vitest.config.ts
Discussion (0)
Ask about usage, or say what you built with itSign in to join the discussion.
No comments yet. Be the first to say what this is good for.
More connectors

Universal provider proxy for OpenAI Codex & Claude Code — use any LLM (Claude, Gemini, Grok, DeepSeek, Ollama…) with Codex CLI, App, SDK, and Claude Code
Git-native persistent memory for AI coding agents. Implements Google OKF v0.2 with sub-300µs in-memory BM25 search, embedded MCP server, and progressive disclosure. Slashes token bloat by 80% with zero external databases or dependencies. Built in pure Go.
Local-first code intelligence graph for MCP and CLI. Builds a persistent map of your codebase so AI coding tools read only what matters, with benchmarked context reductions on reviews and large-repo workflows.
Stop your AI from making things up — it proposes, deterministic tools decide, every claim checked against ground truth with evidence. Grounded facts and context survive resets. Reverse engineering is the proving ground. MCP server + CLI.
20 MB lightweight cross-platform database client for 90+ databases, including MySQL, PostgreSQL, SQLite, Redis, MongoDB, DuckDB, SQL Server, and Dameng. Built-in AI, MCP Server, CLI, desktop and Docker. | 轻量级跨平台数据库管理工具,支持 MySQL、PostgreSQL、SQLite、Redis、MongoDB、达梦等 90+ 数据库,提供桌面端、Docker、CLI、内置 AI 助手和 MCP Server。