MCP: let an AI client drive the app

SeaHi Serial ships a Model Context Protocol server inside the app process. Once a client (Claude Code, Cursor, VS Code, Cline…) connects, the model can open ports, send data, read BLE notifications and run workflows — through the same code paths as the UI, not a separate side channel.

How it runs

Process	In-process with the app, started with it
Transports	Streamable HTTP (POST /mcp) and legacy SSE (GET /sse + POST /messages), same tools and sessions
Binding	Loopback only (127.0.0.1 / ::1 / localhost)
Auth	Token on every request (?token= or Authorization: Bearer); /healthz is the only unauthenticated endpoint and returns {"ok":true}
Toggle	MCP button in the title bar; state lives in ai-config.json

Connect

npx seahi-serial-mcp install          # detect a running instance and write the client config
npx seahi-serial-mcp install --dry-run
npx seahi-serial-mcp status
npx seahi-serial-mcp uninstall

You can also copy the endpoint from the MCP dialog in the app.

Safety model

• Read-only mode: every write tool is rejected with -32007; UI and config stay untouched. Only the user can turn it off.
• Dangerous actions need confirm:true, otherwise the tool returns -32006 with the consequences spelled out.
• Limits: sessions, queue length, request body (1 MiB) and call rate are all bounded; input array/string parameters are checked before the app is touched.

Tools

54 built-in tools in five families: app info, serial semantics, BLE (central), ADB, and UI/log access. The serial happy path:

serial_get_state → serial_select_port → serial_set_baud → serial_open
  → serial_send → serial_get_output → serial_close

The full per-tool reference (arguments and return shapes) is generated from the source: MCP_TOOLS.md (Chinese).