A self-hosted Claude Code that isn't chained to Anthropic

What it does

It’s two things in one repo: a terminal AI assistant in the Claude Code mold, and the Python framework underneath it. The CLI gives you streaming chat, file editing, shell execution, web search, browser automation via Playwright, and a Docker sandbox with named workspaces. The framework side exposes all of that through a single create_deep_agent() call where every capability — subagents, skills, memory, checkpoints, cost tracking — is a boolean toggle.

The interesting bit

The harness treats the boring failure modes as first-class features: stuck-loop detection (including A-B-A-B tool-call patterns), context-limit warnings pushed to the model at 70% token usage before auto-compression kicks in at 90%, and hard budget limits that raise BudgetExceededError. That’s the plumbing most homegrown agents quietly lack.

Key highlights

• Works with any tool-calling model: Anthropic, OpenAI, OpenRouter, Gemini, Ollama for local, or any OpenAI-compatible endpoint
• Persistent memory via MEMORY.md, auto-injected into the system prompt, with a /improve command that proposes updates from past sessions
• Multi-agent teams with shared TODO lists, claiming, and a peer-to-peer message bus
• Checkpoints let you rewind or fork a session to explore alternative approaches
• Headless run mode with JSON output for CI and benchmarks, plus a Harbor adapter for Terminal Bench
• Claims 100% test coverage and strict Pyright/Mypy typing throughout

Caveats

• The feature list is long, and the README doesn’t say how much of it is battle-tested versus recently shipped — several headline features arrived in the v0.3.x series within weeks of each other.
• Docker sandbox and browser automation are opt-in extras, so the out-of-the-box experience is narrower than the table suggests.

Verdict

Worth a look if you want Claude Code ergonomics with model freedom, or a typed Python base for building your own deep agents. If you’re happy with the vendor’s own CLI and don’t need the framework layer, there’s less here for you.