Lohnt sich dein eigenes Composio?

Build a local MCP (Model Context Protocol) server in TypeScript that gives an AI agent working tools for my own SaaS accounts, using personal access tokens. No hosted service, no accounts, no telemetry.

Stack, non-negotiable:
- Node 20+, TypeScript, pnpm
- @modelcontextprotocol/sdk over stdio transport
- zod for tool input schemas
- undici fetch for HTTP, no per-provider SDKs

Structure:
- src/index.ts: server bootstrap, registers every tool module
- src/providers/gmail.ts, slack.ts, github.ts, notion.ts
- src/http.ts: one shared request helper with timeout, retry on 429 and 5xx with exponential backoff, and error normalization
- .env.example listing every token var, .env gitignored, fail loudly at startup if a token for an enabled provider is missing

Tools to implement, keep them small and honest:
- gmail_search_messages, gmail_get_message, gmail_send_message
- slack_list_channels, slack_post_message, slack_search_messages
- github_list_issues, github_create_issue, github_comment_issue
- notion_search, notion_get_page, notion_append_block
Each tool: zod-validated input, a one-line description an LLM can act on, and output trimmed to useful fields only, never raw API dumps.

In scope: static token auth, a PROVIDERS env var to enable a subset, structured logging to stderr only (stdout is the MCP channel), a vitest suite that mocks HTTP and covers happy path plus 401 and 429 handling.
Out of scope: OAuth flows, multi-user credential storage, a web UI, a database. Say so in the README.

Deliver a README with setup steps, token scope requirements per provider, and the exact JSON block to paste into a Claude Desktop or Cursor MCP config.