WorkBuddy Codex Hub MCP v0.1

A local MCP server for coordinating two coding agents through a shared SQLite task pool and an explicit review workflow.

This project is an early v0.1 implementation. It provides a runnable MCP server, a SQLite schema, agent prompt templates, examples, and a smoke test that exercises the core tool flow. It does not yet provide autonomous scheduling, a web dashboard, hosted execution, or production hardening.

What It Does

• Lets agents query open tasks from a shared SQLite database.

• Lets one agent claim and submit work for a task.

• Requires review by a different registered agent.

• Records review decisions and scores.

• Stops repeated change-request loops after max_iterations.

Related MCP server: coordinaut

MCP Tools

Quick Start

npm install
npm test
npm start

The default database path is ./project_hub.db. Override it with:

HUB_DB_PATH=/path/to/project_hub.db npm start

To create sample tasks:

sqlite3 project_hub.db < examples/sample_tasks.sql

Agent Setup

Copy the relevant prompt file into each agent:

• SKILL_workbuddy.md

• SKILL_codex.md

Both agents must connect to this MCP server and use the same SQLite database path.

Verification

npm test

The smoke test starts the compiled MCP server and checks:

• tool listing

• normal query, claim, submit, review, approve flow

• unknown agent rejection

• double-claim rejection

• self-review rejection

• wrong-reviewer rejection

• max-iteration rejection

Publishing

Before publishing, update repository metadata for your actual GitHub location and run:

npm test
npm audit --omit=dev --audit-level=low

This repository includes:

• smithery.yaml for stdio startup configuration.

• Dockerfile for registry or hosted builds.

• LICENSE with MIT terms.

See PUBLISH_TO_SMITHERY.md for a conservative publishing checklist.

Current Limits

• No automatic task scheduler.

• No concurrent write locking beyond SQLite and task status checks.

• No web UI.

• No hosted service.

• No claim that this is unique in the market.

• No claim that registry installs or GitHub stars are predictable.

License

MIT. See LICENSE.