Local Agent Workbench

Local Agent Workbench

A local AI workbench where Claude, Codex and local Qwen work together with durable state, while an optional affordable consultant provides bounded advice.

Local Agent Workbench is an open source macOS application for running real project work through multiple AI runtimes from one persistent app. It is not another chat wrapper: every task has durable state, an isolated worktree, an explainable route, a review bundle and an explicit promotion gate.

Product in one minute

The app opens at login and exposes two native windows:

1. **AI Workspace:** project chat, `Auto / Claude / Codex / Local`, plan, tool stream, files, diff, tests, approvals and provider health.

2. **Planner:** calendar, deadlines, job queue, `Needs you`, watcher inbox, notifications and the audit ledger.

The default flow is:

task → local policy check → plan → isolated worktree → execute
     → verify → review → human approval → apply locally → remember

What makes it different

1. Native Claude Code and Codex runtimes use the user's subscriptions; this project does not turn them into paid API calls.

2. OpenRouter is optional and restricted to `deepseek/deepseek-v4-flash-0731` as an untrusted consultant. It cannot execute tools, write files or approve actions.

3. A local Qwen model suggests classifications and handles private or repetitive work; deterministic backend policy owns routing and fails closed when cloud use is forbidden.

4. Switching runtimes creates a checked handoff package instead of pretending providers share hidden context.

5. `projmem` MCP is the durable project memory engine; SQLite owns operational state only.

6. Agents can write freely inside isolated worktrees, but cannot promote, push, deploy, send or spend without the appropriate gate.

7. No account, no hosted relay and no telemetry by default.

Source and local state

This repository contains only public, reproducible source. Personal hosting state lives outside the repo