AGENTS.md
+11 −4
https://raw.githubusercontent.com/getsentry/sentry-mcp/HEAD/AGENTS.md
− - **MCP tool count**: Target ≤20, hard limit 25 (AI agents have limited tool slots).− - docs/operations/stdio-auth.md — Device code flow, token caching, client ID architecture− - docs/operations/github-actions.md — MCP deployment workflows− - docs/releases/stdio.md — npm package release+ - **MCP tool count**: `EXPOSE_ALL_TOOLS` lists every tool and relies on harness tool search. The gateway surface it falls back to (`TOP_LEVEL_TOOL_NAMES`) targets ≤20, hard limit 25.+ - **Dependency direction**: `sentry` (cli) → `@sentry/mcp-core` → `@sentry/toolkit-core`, and `@sentry/mcp-cloudflare` → `@sentry/mcp-core`; never the reverse. mcp-core and toolkit-core never import the CLI. Inject CLI-only capabilities (SQLite cache, local-file DSN detection, file uploads) through interfaces defined in mcp-core or toolkit-core. `pnpm run lint` enforces this; see docs/architecture/overview.md#package-dependency-direction.+ - **Package CI**: Put checks for one workspace package in its `package.json` scripts and `sentryCi`, which `test.yml` runs per project. Do not add a workflow for them. See docs/operations/github-actions.md#package-ci-checks.+ │ ├── docs-mcp/ # Private Docs MCP Worker+ │ ├── games-api/ # Public Worker for the sentry games leaderboard+ │ ├── codemod-cli3to4/ # @getsentry/cli3to4 codemod: CLI v3 → v4 (Codemod Registry)+ │ ├── logger/ # Structured logger (@sentry/logger on npm)+ - docs/mcp/harnesses.md — First-class agent harnesses and native tool search support+ - docs/operations/stdio-auth.md — stdio auth: shared CLI login and static tokens+ - docs/operations/github-actions.md — CI matrix, `package.json#sentryCi`, and MCP deployment workflows+ - docs/releases/stdio.md — `@sentry/mcp-server` release and the MCP registry