# AGENTS.md — MCP server (TypeScript)

## Setup
- `npm i` · build `npm run build` · test with the inspector: `npx @modelcontextprotocol/inspector node dist/index.js`

## Design rules
- Use the official `@modelcontextprotocol/sdk`. Register tools with `server.registerTool` and zod input schemas.
- Tool names are verbs in snake_case (`search_issues`, `create_note`). Descriptions say what the tool does, when to use it, and what it returns — the model reads them.
- Keep tools small and composable; return concise, structured text. Paginate or truncate large results and say so in the output.
- Mark side effects clearly: read-only tools vs. tools that write/delete. Destructive tools require explicit parameters (no implicit "delete all").
- Return errors as tool results with `isError: true` and an actionable message instead of throwing.
- stdio transport: never write logs to stdout (it corrupts the protocol) — log to stderr.
- Secrets come from environment variables; never echo them in tool output.

## Done
- `npm run build` passes, every tool exercised in the inspector, README lists tools and required env vars.