PocketPaw
Self-hosted AI agent for macOS, Windows and Linux. Use Ollama offline or Anthropic and OpenAI cloud models, with chat apps and a web dashboard.
PocketPaw is a self-hosted AI agent for people who want a personal assistant that can work with files, write code and use a browser on their own computer or server. It runs on macOS, Windows and Linux, with a web dashboard and access through messaging apps. It is open source under the MIT license.
Ollama lets it run models locally, including Llama, Qwen and Mistral, without an API key or account. Local inference can work fully offline; messaging services and web tasks need an internet connection. PocketPaw also supports Claude Agent SDK and OpenAI. With those cloud providers, the agent runs on your machine while model requests go to the provider. PocketPaw collects no telemetry.
Telegram, Discord, Slack and WhatsApp share the same agent, tools and conversation memory. It also supports Signal, Matrix, Microsoft Teams and Google Chat. Session history and long-term facts help it retain context between conversations, and optional Mem0 support adds semantic memory search.
The Command Center splits a goal into tasks, assigns agents and shows their progress, with approval before work ships. Browser automation uses Playwright to read pages and interact with websites. Skills extend what the agent can do, while optional channel packages let you add the connections you need.
PocketPaw encrypts stored credentials and applies tool policies with deny-lists. Guardian AI uses a second model to review dangerous commands before execution, and an append-only audit log records tool activity.
PocketPaw is in beta, and updates may introduce breaking changes. Docker Compose is another documented deployment option.
