UniGrok · one Grok gateway for every coding agent

Run Grok once on your machine. Point every IDE agent at it. Keep the xAI credential on the server — never in each editor.

UniGrok is a local-first MCP gateway for xAI Grok. Cursor, Claude, VS Code, Codex, Antigravity, and other MCP clients share one Streamable HTTP endpoint:

http://localhost:4765/mcp

It routes across the xAI developer API (metered) and the SuperGrok CLI subscription (when authenticated), and returns the answer under response with route / plane / cost metadata.

Current release: v0.6.0.

---

1. What you get

• Shared @grok / agent tool for every IDE on your laptop
• Server-side credentials (API key and/or CLI OAuth)
• Optional local Core UI for status: http://localhost:4765/ui/
• A habit you can opt into: critique Implementation Plans with Grok before the user sees them

Your other projects do not need UniGrok repo folders, contributor trees, or special mounts.

2. Prerequisites

• Git
• uv
• An MCP-capable IDE
• At least one Grok credential path:
  • xAI API key from console.x.ai, or
  • SuperGrok / Grok CLI device login (subscription)

Docker Desktop is the current packaging path for the shared service. Treat it as how UniGrok runs on your machine — not as “you are a Docker developer.”

3. Run the gateway (one path)

[!WARNING] UniGrok is not published on PyPI. pip install mcp-grok installs an unrelated project; use this GitHub checkout. There is no real npx unigrok server yet — Docker (or uv + compose) is the install.

Fast path (copy/paste):

git clone https://github.com/djtelicloud/grok-mcp-server.git && cd grok-mcp-server
uv run python main.py init
docker compose up --build -d

init creates .env if missing and prints IDE MCP snippets for http://localhost:4765/mcp. After the service is healthy, you can close this repo and work in any project with @grok.

Before checking readiness, complete one credential path:

• xAI API: set the developer key in server .env before starting Compose (or rerun docker compose up -d after editing it).
• CLI subscription: after the image is up, run:

docker compose run --rm grok-cli-auth

Then verify the usable plane:

curl --fail -s http://localhost:4765/readyz

You are ready when /readyz reports "status":"ready". /healthz only proves the process is up; CLI-only installs must authenticate before this readiness check.

4. Connect your IDE (paste this to your agent)

Copy the block for your IDE from the init output, or paste this prompt to your local coding agent:

Configure UniGrok MCP for this machine:
- Streamable HTTP URL: http://localhost:4765/mcp
- Send a stable X-Client-ID header for this IDE (e.g. cursor, claude-code, vscode, codex)
- Never put XAI_API_KEY in IDE MCP settings; credentials stay in UniGrok's server .env
- After connecting, call tools/list and grok_mcp_discover_self
- Prefer the UniGrok agent tool when I say @grok or want a second opinion
- When I ask for a multi-step Implementation Plan, get a UniGrok second opinion
  (agent mode thinking or reasoning) and improve the plan before showing it —
  only if I want that habit; do not silently spend metered API credits
- Do not invent a second MCP port, Forge, or land workflow for ordinary use

Detailed multi-IDE notes: docs/ide-setup.md (public path only; contributor dual-runtime detail lives in CONTRIBUTING).

5. Prove it in 60 seconds

1. Restart the IDE MCP client.
2. Ask: “Call UniGrok discover_self and tell me which credential planes are ready.”
3. Ask for a small plan, then: “Get a UniGrok second opinion on that plan before you show me the Implementation Plan.”

Optional status UI (this machine only): open http://localhost:4765/ui/.

6. Security (short)

1. Keep XAI_API_KEY and CLI OAuth on the UniGrok service — never in IDE configs.
2. API plane is metered; CLI subscription cost is not exposed by the provider — UniGrok only tracks local CLI activity.
3. Stable UniGrok does not browse your project unless you pass workspace_context.
4. Loopback Core UI is for the machine owner; it is not a public multi-user console.
5. Do not paste secrets into chat.

7. Project site (optional)

The public site at https://grokmcp.org is bound to the existing project repository. It is not an idless installer template. For the optional contributor control surface, GitHub App OAuth establishes identity and each privileged request performs a fresh installation-token lookup of repo role; a Sites rollback fallback remains only for the legacy identity binding path. That control plane does not replace local UniGrok MCP on localhost:4765, and it never holds your xAI key.

8. Where docs live

You are…	Use
Installing / connecting an IDE	This README
An agent needing schemas and operations	OKF knowledge bundle (also via discover_self and local /docs/okf/)
Public “smarter defaults” recipes	docs/public-intelligence/
Changing UniGrok itself	CONTRIBUTING.md

Source of truth: this repo’s docs/okf/ + the site OKF.
GitHub Wiki (optional): human-friendly mirror only, generated from OKF (see docs/wiki-okf-mirror.md). Do not hand-edit the wiki as product docs. Empty or stale wiki is not an outage — use OKF.

9. Next steps

I want…	Go here
More IDE examples	docs/ide-setup.md
Agent knowledge (OKF)	https://grokmcp.org/docs/okf/index.md
Architecture	architecture.md
Security reporting	SECURITY.md
Contribute to UniGrok itself	CONTRIBUTING.md
Project site	https://grokmcp.org

Contributors and admins: dual-runtime Forge, landing, Swarm, and insider Console behavior are documented only in CONTRIBUTING.md — not required for ordinary use.

License

MIT