UnClaude

UnClaude is an autonomous coding agent you install with pipx and point at whichever model you like, including local ones with no API costs. It routes each request to a model that fits the job, remembers across sessions, and can keep working in the background on a task queue. It ships with a web dashboard, messaging, a multi-agent swarm, and audit logging of every tool call.

How it works

1. 01

   Each request is scored on length, code, reasoning markers and conversation depth, then sent to the cheapest model tier that can handle it.

2. 02

   A hierarchical memory stores resources, items and categories with salience scoring, so important context surfaces first.

3. 03

   A background daemon takes tasks from the CLI, files, webhooks, cron or git hooks and runs them under capability-based security profiles.

What it does

• Any LLM through LiteLLM, including fully offline local models
• Smart routing with auto, eco, premium and free profiles to match model cost to request difficulty
• Autonomous daemon with a priority task queue and a proactive behaviour file the agent reads live
• Ralph Mode plans, codes, runs your feedback command and fixes its own failures
• Web dashboard, Telegram and WhatsApp messaging, usage budgets and a sandboxed, audited tool layer