Captain Claw is a local AI agent runtime I've been building. Install with pip install captain-claw or Docker, point it at any provider (OpenAI, Anthropic, Gemini, Ollama), and get a persistent multi-session agent with a web UI out of the box. The thing I haven't seen elsewhere: sessions are first-class citizens. You can run session #1 on Claude and session #2 on GPT simultaneously, pass context between them, and resume everything exactly where you left off - backed by SQLite. 29 built-in tools the agent picks automatically — shell, files, web search/fetch, PDF/DOCX/XLSX/PPTX extraction, image gen/OCR/vision (DALL-E 3), Gmail (read), Google Calendar, Google Drive, email sending, TTS, todo, contacts, cron, a relational datastore with web dashboard, deep memory via Typesense, playbook distillation, personality profiles, BotPort agent-to-agent routing, and even Termux for Android (camera, GPS, torch, battery). Orchestrator mode decomposes complex tasks into a parallel DAG across separate sessions with real-time progress monitoring. Also ships with an OpenAI-compatible /v1/chat/completions proxy so you can point existing tools at it.

Terminal demo: https://asciinema.org/a/814073

Video: https://www.youtube.com/watch?v=4g_aA_WnEaw

GitHub: https://github.com/kstevica/captain-claw

Happy to answer questions about the architecture or design decisions.