Introducing PandaClip🐼
I've spent the last few months building something I wanted for my own AI workflows.
PandaClip is an open-source, local-first MCP toolbox for AI agents with a live desktop lens that lets you watch everything your agents do as it happens.
Instead of piecing together dozens of utilities, PandaClip provides a shared local workspace with over 40 deterministic tools across four systems:
📋 Clipboard — clipboard history, snippets, channels, TTL-based cleanup, and built-in secret screening.
⚡ Cache — a namespaced TTL cache so agents stop repeating expensive lookups.
🎋 Bamboo — organize existing files with tags, notes, and metadata without moving or renaming them.
🌱 Garden — a knowledge graph where agents can plant facts, connect them with typed relationships, track revisions, and build persistent knowledge over time.
The feature I'm most excited about is the Lens.
Leave it open on a second monitor and watch your agents work in real time. Every clipboard update, memory write, knowledge graph change, project update, and note edit appears instantly in a live activity feed. Click any event to inspect its contents, or keep the frameless hover overlay pinned while you work.
Everything is local-first.
• No cloud services
• No remote database
• No network dependency
• No background daemons
• The MCP server is the only writer, while the Lens is a read-only observer designed to safely monitor activity without interfering with your agents.
The idea was inspired by Anthropic's Jacobian Lens.
Jacobian Lens visualizes what's happening inside a language model. PandaClip brings that idea one level higher—it visualizes your agents' working state as they interact with tools, memory, projects, and knowledge.
PandaClip also pairs naturally with magpie-search.
magpie-search is responsible for finding information across conversations, files, and the web. PandaClip provides the structured workspace where agents cache results, save reusable information, and grow a persistent knowledge graph. They're intentionally independent projects that complement one another.
Every feature shown has been tested, and every claim in this post reflects the current implementation.
The project is open source under Apache-2.0.
I'd love feedback from anyone building AI agents, MCP servers, or local-first tooling.
🐼 Welcome to PandaClip.