I build small, offline-first tools for running AI agents and MCP (Model Context Protocol) servers safely on real machines. Most of my work comes from operating a multi-PC agent setup day to day: when something breaks there, it becomes a tool here.
What I work on
- MCP tooling β lint, diff, audit and test MCP tool contracts before they reach production.
- Agent safety β durable locks, replay guards and tamper-evident audit logs for autonomous agents.
- Local search β dependency-free indexing for personal knowledge bases.
Start here
| Project | What it does |
|---|---|
| joons-ecosystem | Map of the JOONS tools and how they fit together |
| joons-mcp-config-doctor | Offline linter and secret redactor for MCP client configs |
| joons-search-engine | Offline, dependency-free search for local documents |
| joons-agent-guard | SQLite-backed locks, circuit breaker and replay guard for agents |
Open source
I also send focused fixes upstream β tests, docs and small bugs β mostly to Python and TypeScript AI tooling. Selected merged contributions:
- Tencent/teamai-cli β recall ranks confidence by upvote time, not creation time
- docker/docs β restored a broken Docker Desktop docs URL
- dapr/dapr-agents β least-privilege tokens for every CI workflow
- py-pdf/pdfly β
--output-dirforextract-images, shipped in the 0.7.0 release - obsidianmd/obsidian-help β PDF export docs for Obsidian mobile 1.15
- eugeniughelbur/obsidian-second-brain β vault search exposed to scripts without an agent loop
- calibrain/shelfmark β HTTP proxy applied to Open Library and Prowlarr requests
- doobidoo/mcp-memory-service β dashboard loads a single three.js core
- mxsm/rocketmq-rust β Proxy documentation corrections
All of them: see my merged pull requests.
All tools are early (0.x). Each README states its scope and limits honestly.



