Adopt AGENTS.md (with CLAUDE.md pointer) modeled on the inspector repo

Goal

Adopt the agent-guidance pattern from the inspector repo — specifically the AGENTS.md on v2/main, which is the proper reference (the main-branch version is a thin early draft). A single AGENTS.md dev guide that all coding agents read, with CLAUDE.md reduced to a one-line pointer:

@./AGENTS.md

Rules become tool-agnostic — Claude Code, Copilot, Cursor, etc. consume the same document — and nothing is maintained twice.

Starting point

This repo already has a substantial CLAUDE.md (monorepo structure, build/test commands per language, code style, CI notes, contributing guidelines). So this is a migration + extension: move that content into AGENTS.md, reorganize it into the v2/main reference's shape, extend with the sections below, then replace CLAUDE.md with the pointer.

Content plan — mapped from the v2/main reference

Project Structure — the annotated src/ tree (7 servers, 4 TS / 3 Py) with each server's package name and registry. (Adapts inspector's annotated tree.)

Development setup / Build & test commands — largely exists in current CLAUDE.md: npm workspaces for TS (Node 22, vitest + coverage-v8 required for new tests), uv/hatchling for Python (pytest, pyright, ruff).

Repository & Project Board — servers-specific equivalent of inspector's section:

• Repo: https://github.com/modelcontextprotocol/servers.git — base branch main
• Board: Servers V2 board (project 43 — the single board; all work goes here)
• Include the gh recipes with the stable-ID table, exactly as the reference does:

Thing	ID
Project node ID	PVT_kwDOCt2Azc4BcZgq
Status field ID	PVTSSF_lADOCt2Azc4BcZgqzhXCpm0

Status	Option ID
Todo	f75ad846
In Progress	98236657
In Review	47fc9ee4
Done	25b87ca0

Issue-driven Work Style — adapt the reference's rules to this repo:

• An issue isn't "created" until it is labeled v2 and on board 43 and has a Status — three distinct steps
• Only issues go on the board, never PRs (PRs tracked via their linked issue's card)
• Every board item is a real issue (no draft cards); check for duplicates before creating; assign the creator
• Status flow: Todo → In Progress (work begins, feature branch) → In Review (PR open) → Done (merge)
• PR body's first line: Closes #<ISSUE>. Simpler than inspector here: our PRs target main (the default branch), so closing keywords do auto-close the issue on merge — but the board Status still needs to be moved to Done manually (auto-close doesn't touch the project board)
• New tasks discovered mid-work → new issues on the board

Maintenance Rules — keep docs in sync (README, per-server READMEs, RELEASING.md, and AGENTS.md itself when implementation/testing rules change). Reference RELEASING.md for the release process rather than duplicating it.

Always test new or modified code — vitest for TS (with @vitest/coverage-v8), pytest for Python; tests required for new/changed code. (We don't inherit inspector's 90% per-file gate on day one — set expectations honestly and tighten later if wanted.)

Responding to Code Reviews — copy the reference's etiquette verbatim: not every suggestion must be implemented; divergence or rejection is fine with a reason; respond to each review comment with what was done or why not.

Lint-fixed, Formatted code + validate (decided — in scope for this issue):

• Add Prettier (root config, format / format:check scripts) and ESLint (flat config) shared across the TS workspaces
• Root npm run validate = format:check → lint → build → test, chaining the per-workspace validations exactly like the reference's per-client chain
• The rules, per the reference: always npm run format before committing; npm run validate before pushing
• CI (typescript.yml) runs validate so the checks gate PRs
• Python's equivalent one-liner documented per server (uv run pyright && uv run ruff check . && uv run pytest)

TypeScript instructions — adopt the reference's rules (never any; never suppress compiler/linter errors via config; leverage inference, generics, type guards) plus servers-specific idioms: ES modules with .js import suffixes, Zod schemas for tool input validation, naming conventions (camelCase/PascalCase/UPPER_CASE/kebab-case, verb-first tool names), MCP patterns (registerTools/registerResources/registerPrompts, tool annotations readOnlyHint/idempotentHint/destructiveHint, stdio + Streamable HTTP transports).

Python instructions — the counterpart section inspector doesn't need: type hints enforced by pyright, ruff clean, async/await idioms (pytest-asyncio), per-server module layout.

Contribution boundaries — carried from current CLAUDE.md: accepted / selective / not-accepted, no new server implementations (registry instead), PR checklist expectations.

(Inspector's React/Mantine section and web-auth-token section have no equivalent here — omitted.)

Acceptance

• AGENTS.md at repo root, organized per the v2/main reference, containing the merged + extended guidance
• CLAUDE.md replaced with the one-line @./AGENTS.md pointer
• Prettier + ESLint configured at the root, shared across TS workspaces
• Root npm run validate (format:check + lint + build + test) passes on a clean checkout
• CI (typescript.yml) runs validate so the checks gate PRs
• Board section's gh recipes verified working (IDs above are live — they were used to place this very issue)
• Verified Claude Code picks up the guidance through the pointer