Why
Why I built it
Local MCPs run in different ways. I wanted remote AI clients to reach them while keeping each service’s auth, process, and data isolated—without rebuilding exposure plumbing for every new tool.
What I built
What I built
- Relay: per-service OAuth, HTTP sessions, auto-start, and process cleanup via Windows Job Objects, with DPAPI-protected secrets; the relay handles transport, auth, and lifecycle only.
- Camera MCP: 15 tools pairing local Whisper (CUDA) STT, OpenCV person detection, and Windows TTS with MCP-native image and audio payloads.
- Reverse-engineered the camera pan/tilt protocol and wired it into a custom go2rtc build.
Engineering
Problems I worked through
- Shared transport layer while each service retains independent OAuth and sessions; Windows Job Objects keep stdio process trees clean on shutdown.
- Tools return actionable recovery hints and distinguish live feeds from replays so agents react to real state instead of blindly retrying.
- Layered checks across unit tests, the local Windows bridge, and live remote MCP clients keep new local tools reliable end to end.
Role & AI
How I split the work with AI
I set the tool contracts, permission boundaries, and verification checks while coding agents handle implementation; home footage and config never leave my machine.
Outcome
Outcome
The shared relay connects Coding, Anki, and camera services in daily use, letting remote AI clients safely drive local tools and camera controls.