Skip to main content
Glama
Sign In
CodexMCP

CodexMCP

by JohanFJ101
OverviewSchemaSourceRelated ServersDiscussionsAdmin
JavaScript
Local

CodexMCP

Personal MCP server that exposes a query_offload tool to Claude. Lets Claude delegate mechanical, output-heavy tasks (boilerplate, type generation, summarisation) to OpenAI while keeping reasoning in-context.

Setup

1. Install dependencies

npm install

2. Make codexmcp available globally

npm link

This creates a global codexmcp binary pointing at bin/codexmcp.js.

3. Configure environment

cp .env.example .env

Edit .env and fill in:

Variable

Description

TUNNEL_URL

Your Tailscale Funnel URL (e.g. https://t480.tail1234.ts.net)

JWT_SECRET

Any random 32+ character string

PORT

Port to listen on (default 3000)

Inference runs through the Codex CLI, not an OpenAI API key. Install it and run codex login (ChatGPT subscription) on the host machine before starting the server — query_offload shells out to codex exec.

4. Tailscale Funnel (T480)

Tailscale Funnel exposes a local port on a stable public HTTPS URL — no dynamic DNS, no port forwarding.

# On the T480, run once (survives reboots if you enable it):
tailscale funnel 3000

Get your permanent URL:

tailscale funnel status
# Example output: https://t480.tail1234.ts.net -> localhost:3000

Copy that URL into TUNNEL_URL in your .env.

5. Start the server

codexmcp

Expected output:

codexmcp running on http://0.0.0.0:3000
MCP endpoint (add to claude.ai): https://t480.tail1234.ts.net/mcp

6. Verify

curl http://localhost:3000/health
# {"status":"ok","server":"codexmcp","version":"0.0.1"}

7. Connect to claude.ai

  1. Go to claude.ai → Settings → Connectors → Add custom connector

  2. Paste your TUNNEL_URL/mcp (e.g. https://t480.tail1234.ts.net/mcp)

  3. Complete the OAuth flow — it auto-approves (single-user server)

8. Create a Claude Project

  1. Create a new Project in claude.ai

  2. Add the codexmcp connector to it

  3. Paste the system prompt below into Project Instructions

Related MCP server: Houtini LM

Permanent deployment with systemd (T480)

Copy the repo to the T480, install, link, and deploy the service:

# On T480
git clone <repo> ~/codexmcp
cd ~/codexmcp
npm install
npm link

# Copy and enable the service
sudo cp offload-mcp.service /etc/systemd/system/
sudo systemctl daemon-reload
sudo systemctl enable offload-mcp
sudo systemctl start offload-mcp

# Check status
sudo systemctl status offload-mcp
journalctl -u offload-mcp -f

The ExecStart in offload-mcp.service points to /usr/local/bin/codexmcp (where npm link installs the global binary on Linux). Adjust if your npm prefix is different (npm prefix -g to check).

Claude Project system prompt

Paste this verbatim into the Project Instructions field:

## Behavior
- Never glaze. Be direct, honest, no sycophancy.
- Make no assumptions. Use the MCQ tool to clarify before proceeding. If skipped, pick the best option yourself.
- Format all responses carefully in clean markdown.
- Think before responding unless the query is trivial.
- When asked for opinions, be unbiased — say what is true, not what I want to hear.

## Task Routing
For every response, mentally split the task first:
1. Pure mechanical output (boilerplate, type generation, bulk code, summarizing large text, repetitive patterns) → call query_offload, return the result directly.
2. Mixed task (reasoning + output) → handle the reasoning yourself, call query_offload for the output part.
3. Pure reasoning (debugging, architecture, review, explanation) → answer directly, no tool call.
Call query_offload automatically. Never ask if you should use it.

Development

npm run dev   # node --watch src/server.js

This server cannot be deployed

Claude Code, inside the Godot 4 editorFlockbay puts an AI agent in the real Godot 4 editor: it writes GDScript, runs your game and fixes what breaks. Free, or bring your own Claude Code or Codex.flockbay.comAd

Maintenance

ActivityInactive
ResponsivenessNo issues

Resources

Related MCP Connectors

Related MCP Servers

  • Houtini LM

    Kerstin59718
    F
    license
    Not graded
    quality
    A
    maintenance
    Enables Claude Code or any MCP client to delegate bounded tasks such as code review, boilerplate, and documentation to local or cloud OpenAI-compatible LLMs. This reduces token usage by keeping large source files out of the main model's context while allowing results to be reviewed by the orchestrator.
    -

  • Houtini LM

    Pavelj43
    A
    license
    Not graded
    quality
    A
    maintenance
    Enables MCP clients such as Claude Code to delegate bounded coding and review tasks to local or cloud OpenAI-compatible LLMs, reducing token costs and providing a second model opinion. The main orchestrator retains planning and tool use while offloading drafting, code review, test stubs, documentation, and embeddings.
    MIT

  • Houtini LM

    JeanShah47544
    A
    license
    Not graded
    quality
    A
    maintenance
    Enables MCP clients such as Claude Code to delegate bounded tasks like code review, boilerplate generation, commit messages, format conversion, and embeddings to local or cloud OpenAI-compatible LLMs, reducing token costs while keeping the main model in control.
    MIT

  • Houtini LM

    ThuSyed
    A
    license
    Not graded
    quality
    A
    maintenance
    Enables Claude and other MCP clients to hand off bounded work — code review, test stubs, boilerplate, commit messages, docs, format conversion, embeddings — to local or cloud OpenAI-compatible models while the main assistant keeps planning and QA. Keeps large source files out of the orchestrator's context window, cutting token costs and providing a cheap second opinion on code.
    MIT