🦞 Pynchy (pronounced "Pinchy") — A personal AI assistant like OpenClaw done right. Security first, modular, written in Python.
Everyone is writing their own AI assistant. Why write another one? The biggest reason is that I wanted something written in Python, because that's what I'm most comfortable with.
- ZeroClaw looks great actually, but I don't know how to write in Rust.
- Happy looks great, but ultimately is a remote terminal to Claude Code. I want to add my own security features. Also, I am not fluent in TypeScript.
- NanoClaw is a too minimalist.
- OpenClaw is a massive pile of overcooked spaghetti code. Ain't no way I'm running that security nightmare on my machine.
- pi mono is a less crazy project, which actually OpenClaw built on top of. It doesn't have the security features that I want.
- Agents run in containers, providing process, filesystem, and network isolation.
- Built-in plugins ship with the monorepo; third-party plugins are discoverable via Python entry points.
- Uses LiteLLM as the LLM gateway, providing a bunch of features out of the box:
- Automatic load balancing across APIs, to soak up your various allowances from different providers.
- Access to 100+ LLM providers
- Cost tracking and budget management.
- Rate limiting
- MCP gateway — centralized management of external MCP tool servers with per-workspace access control, on-demand Docker lifecycle, and config-driven setup.
- (see the LiteLLM docs for more details)
- Customizable; eight types of plugins are supported — agent cores, skills, channels, service handlers, container runtimes, workspaces, observers, and tunnels.
- Persistent memory with BM25-ranked full-text search — agents save and recall facts across sessions.
- Reoccurring tasks can be scheduled to run at a specific time or interval.
- (work in progress) policy groups to prevent lethal trifecta prompt injection attacks.
Built-in plugins provide integrations with external services. All integrations are pluggable — see plugin authoring to add your own.
| Integration | What it does |
|---|---|
| Messaging channel via linked device | |
| Slack | Messaging channel with browser-based token extraction |
| X (Twitter) | Post, like, reply, retweet, and quote via browser automation |
| CalDAV | Calendar access (Nextcloud, etc.) — list, create, delete events |
| Jupyter Notebooks | Per-workspace notebook server with MCP tools |
| Google Drive | File access via OAuth2 MCP server |
See the installation guide to get started.
Full documentation at pynchy.ricardodecal.com.
| Section | What it covers |
|---|---|
| Usage | Day-to-day operation, groups, scheduled tasks |
| Plugin authoring | Writing plugins: channels, skills, MCP servers |
| Architecture & Design | Container isolation, message routing, IPC, security |
| Contributing | How to contribute — plugins, fixes, docs, and more |
What messaging channels are supported? WhatsApp and Slack have first-party plugins. Channels are pluggable — write a plugin to add new ones.
Why Apple Container instead of Docker? On macOS, Apple Container is lightweight and optimized for Apple silicon. Docker works too and is used as a fallback. On Linux, Docker is the only option.
Is this secure? Agents run in containers, not behind application-level permission checks. They can only access explicitly mounted directories. See the security model for details.
How do I debug issues? Ask Pynchy. "Why isn't the scheduler running?" "What's in the recent logs?" That's the AI-native approach.
Huge thanks to NanoClaw. This project started as a Python port of that project.
MIT