PocketPaw
A self-hosted personal AI assistant for your computer, browser dashboard, and supported chat channels.
Per-dimension scores and reasoning
Evidence shows security features (Guardian AI, injection scanning, tool policy, plan mode, audit log), but implementation details are not provided. Least privilege, user confirmation, data flow transparency are only partially addressed. Sensitive data handling mentions API key encryption but not the mechanism. Dependency security has version constraints but no vulnerability scanning mentioned. External effects have tool policy but no explicit default restrictions. Rollback mechanism not mentioned. Source attribution has author info but unverified.
Self-consistency is decent; README and pyproject versions match (0.4.18), but inconsistencies exist: README claims '50+ tools' but pyproject doesn't list all; README mentions '9+ channels' but pyproject lists only some. Dependency availability has version ranges but no lock file. Failure messages have health check hints but no detailed error handling docs.
Target audience is clear (individual users, developers), scenarios diverse (chat, automation, research). Capability boundaries have security policies but not detailed. Trigger precision has tool policy but no explicit trigger conditions. Environment fit supports multiple platforms, Docker, WSL2, but no system requirements details.
Information architecture is clear with README, docs links, architecture diagrams. Install notes are detailed covering multiple platforms. Naming stable but version updates frequent (0.4.18). Examples and FAQ have example dialogues but no FAQ. Known limitations only mention Beta status. License MIT clear. Versioning changelog not provided but pyproject has update comments. Maintenance responsibility has team info but not explicit.
Output usability has example dialogues but no actual output format. Marginal value high (local-first, multi-backend) but no comparison with other products. Cost-benefit has free option (Ollama) but no pricing info.
Claim traceability partially has docs links but no specific evidence. Cross-source corroboration has PyPI, GitHub links but not independently verified. Fact-inference separation not explicit.
- Implementation details of security features are not provided; code review needed to confirm Guardian AI etc. actually work.
- Dependencies are not pinned, posing supply chain risk.
- Features claimed in README (e.g., 50+ tools) are not fully reflected in pyproject; needs verification.
- Rollback mechanism not mentioned; may not recover from failed upgrades.
- Publisher identity unverified; proceed with caution.
What does this agent do, and when should you use it?
PocketPaw is a self-hosted personal AI assistant with a Python backend, native desktop client, and browser dashboard. Messages enter an event-driven message bus, where AgentLoop routes them to the configured AgentBackend. It supports Discord, Slack, WhatsApp, Telegram, and other documented channels, with backends for Claude Agent SDK, OpenAI Agents, Google ADK, Codex CLI, OpenCode, and Copilot SDK. The desktop client in client/ is a Tauri 2.0 and SvelteKit application that connects to the Python backend over REST and WebSocket. The dashboard runs locally, while Docker deployments expose agent-created files through the host's workspace/ directory.
Channels publish messages to an event-driven message bus, and AgentLoop picks them up for the configured AgentBackend. The documented tool set includes browser access, web search, image generation, voice, OCR, research, delegation, skills, and integrations; results are returned through the dashboard or connected chat channel. PocketPaw stores long-term facts and session history, with smart compaction and optional Mem0 semantic search. Guardian AI can review every tool call before execution, alongside injection scanning, configurable tool policies, Plan Mode, --security-audit, and a self-audit daemon with append-only audit logs.
- An individual who wants a personal assistant to run on their own machine and converse through a browser or Telegram.
- A team or operator who needs to connect an assistant to Discord, Slack, WhatsApp, or Telegram.
- A user who wants a local inference option through Ollama for fully offline operation.
- A product or business user who wants multiple agents to work on research such as a competitor analysis report.
- A security-conscious user who wants tool calls reviewed through Guardian AI, Plan Mode, and audit logging.
What are this agent's strengths and limitations?
- One AgentBackend protocol spans six documented backend options: Claude Agent SDK, OpenAI Agents, Google ADK, Codex CLI, OpenCode, and Copilot SDK.
- It combines a local web dashboard, a native Tauri desktop client, and more than nine documented communication channels.
- Ollama provides a documented path for local, fully offline operation.
- Its documented security stack includes Guardian AI review of tool calls, injection scanning, tool policy controls, Plan Mode, and append-only audit logs.
- The project is in beta and explicitly warns that versions may introduce breaking changes.
- A first run without at least one configured model provider reports UNHEALTHY and leaves AI features disabled.
- Backend capabilities differ: the README marks MCP support only for Claude Agent SDK, Google ADK, and Codex CLI.
- Browser automation and shell tools are documented as working best under WSL2 on Windows; native Windows coverage is specifically stated for the dashboard and LLM chat.
How do you install or deploy this agent?
For terminal installation, install Python 3.11+ and pip, then run pip install pocketpaw && pocketpaw. Open http://localhost:8888 and configure at least one provider key in Settings > API Keys: Anthropic, OpenAI, or Google Gemini. Equivalent environment variables are POCKETPAW_ANTHROPIC_API_KEY, POCKETPAW_OPENAI_API_KEY, and POCKETPAW_GOOGLE_API_KEY. For Docker, run git clone https://github.com/pocketpaw/pocketpaw.git && cd pocketpaw, then cp .env.example .env and docker compose up -d. Native desktop installers are documented for Windows, macOS, and Linux.
How do you use this agent?
Run pocketpaw, then open http://localhost:8888. Save at least one provider key under Settings > API Keys and, if needed, select an agent backend with POCKETPAW_AGENT_BACKEND, such as claude_agent_sdk, openai_agents, or google_adk. Chat in the dashboard or connect Discord, Slack, WhatsApp, or Telegram from it. In Docker, retrieve the access token with docker exec pocketpaw cat /home/pocketpaw/.pocketpaw/access_token; for host-based Ollama, set POCKETPAW_OLLAMA_HOST=http://host.docker.internal:11434 in .env.
How does this agent compare with similar options?
Claude Agent SDK is the default backend; OpenAI Agents, Google ADK, Codex CLI, OpenCode, and Copilot SDK are documented configurable alternatives. Ollama is the local and offline option, while Anthropic, OpenAI, and Google Gemini require their respective API keys.
FAQ
Does PocketPaw require a subscription?
Why does the dashboard show UNHEALTHY after installation?
Where are settings and generated files stored?
~/.pocketpaw/config.json, and API keys are encrypted at rest. In Docker deployments, agent-created files appear in the host ./workspace/ directory.Can I require approval or review before tool actions run?
--security-audit, and a self-audit daemon.