shaughv-health-mcp

A private, single-user remote MCP server on Vercel for Emmett. His trusted LLM assistants — Claude Code / Claude Desktop, claude.ai (web + mobile connectors), and ChatGPT connectors — can read his Google Health data ("How many steps today?", "How did I sleep?", "Why am I tired?", "What did I eat yesterday?") and write the things a wearable can't track itself: nutrition, hydration, and body measurements.

It is a thin, typed, authenticated data adapter over the Google Health API (health.googleapis.com/v4). The LLM does the reasoning; this server returns accurate data with timestamps, units, and freshness metadata on every response. It is not a health warehouse, an analytics product, or anything that diagnoses — there are no medical claims anywhere in the surface.

docs/PLAN.md is the build plan and the source of truth for architecture, the four auth layers, the database schema, the tool surface, the watchouts, and the E2E verification bar. Accepted architecture decisions under docs/adr/ govern their specific scope; ADR-0002 fixes the Emmett-only audience while ADR-0001 records the private/unverified posture. This README is the human front door. Live phase/task status is on the .tasks/ board — see Task board.

Table of contents

• What it is / what it is not

• Who it's for and how it's used

• Architecture

• The four auth layers (never conflate)

• Tech stack

• Current status

• Repository structure

• Setup guide

  • 1. Accounts and services

  • 2. Google OAuth redirect URIs

  • 3. Environment variables

  • 4. Local development

  • 5. Deploy

• Connecting an MCP client

• MCP surface

• Security posture

• Troubleshooting

• Task board

• Further reading

Related MCP server: google-health-mcp-server

What it is / what it is not

The device data path is never live — a Fitbit Air syncs to the Fitbit app, which pushes to the Google Health cloud, which this server reads on demand. Every response therefore carries freshness metadata (retrievedAt, latestDataTime?, isPossiblyStale, note) so the LLM can distinguish "no data synced yet" from "nothing happened."

Who it's for and how it's used

Audience: Emmett, through either approved alias, and the LLM assistants he authorizes. Nobody else can sign in: the exact approved identities are fixed by ADR-0002, and enforced through ALLOWED_GOOGLE_EMAILS.

Intent — how Emmett uses it. He asks his assistants natural questions and expects grounded answers with real numbers, timestamps, and freshness notes:

• "How many steps do I have today? What's left to hit my goal?"

• "How did I sleep last night?" / "Why am I tired?"

• "Why is my heart rate so high?" (latest synced HR, explicitly not live)

• "What did I eat yesterday?" / "What's my exercise looking like this week?"

• "Log a snack: Greek yogurt, 150 cal, 15 g protein." / "Edit that to 180 cal." / "Delete it."

• "Log 16 oz of water." / "Update my weight to X."

Writes are the reason the server exists at all beyond read-only: a Fitbit Air does not track food, water, or manual measurements, so the LLM logs them through validated, audit-logged write tools. Every mutation is explicit-input-only (never inferred) and leaves an audit row.

Future directions (parked on the board; see Task board):

• #api — a plain REST API surface (bearer/PAT-authenticated) over the same health services, so Emmett's own scripts/apps can call the data without an OAuth dance. Feasibility analyzed: feasible and additive on the current stack.

• #rlw — closed by ADR-0003: retain Vercel Node 24 + Fluid in iad1; reconsider Edge/Workers/Railway only on measured triggers.

• #w11 — v1.1 webhooks: implemented endpoint/signature/ledger/inbox; production subscriber registration and a real sync are the final live gates.

Architecture

                    ┌─────────────┐   Bluetooth sync   ┌──────────────┐
   Fitbit Air ─────▶│  Fitbit app  │ ─────────────────▶│ Google Health │
   (wearable)       │  (phone)     │                    │    cloud     │
                    └─────────────┘                     └──────┬───────┘
                                                               │  Google Health API
                                                               │  health.googleapis.com/v4
                                                               │  ▲ DPoP-bound refresh token
                                                               │  │  (AES-256-GCM
                                                               │  │   encrypted in Neon)
                                                               ▼  │
   ┌───────────────────────────┐   OAuth 2.1 + DCR    ┌────────┴──────────────────────┐
   │  LLM client               │ ───────────────────▶ │  Next.js on Vercel            │
   │  Claude Code / Desktop    │                       │  (health.emmetts.dev)         │
   │  claude.ai web + mobile   │ ◀─────────────────── │                               │
   │  ChatGPT connector        │   MCP tools/resources │  ┌─────────────────────────┐  │
   └───────────────────────────┘                       │  │ MCP SDK v2   → /api/mcp │  │
                                                        │  │ better-auth  → OAuth AS │  │
                                                        │  │ Google Health client    │  │
                                                        │  └───────────┬─────────────┘  │
                                                        └──────────────┼────────────────┘
                                                                       │ Drizzle ORM
                                                                       ▼
                                                             ┌──────────────────┐
                                                             │  Neon Postgres    │
                                                             │  encrypted tokens │
                                                             │  audit log        │
                                                             │  webhook tables   │
                                                             │  (dormant, v1.1)  │
                                                             └──────────────────┘

Request path in one line: an LLM client completes OAuth 2.1 + S256 PKCE (with Dynamic Client Registration) against this app's own authorization server, gets a one-hour RS256 JWT bound to /api/mcp, and presents it on every request. The MCP route verifies that JWT locally, then a tool handler fetches Google Health with the decrypted-on-demand Google credential, bounds/normalizes the payload, and returns it with freshness metadata.

The four auth layers (never conflate)

These are four separate things. Confusing them is the number-one source of bugs here.

Layers 3 and 4 use the same Google OAuth client ID but separate flows and scopes. Layer 4's login uses basic openid email profile scopes only; layer 3 requests the nine googlehealth.* scopes. The health-flow refresh token is the one that gates data access.

Tech stack

Current status

Milestone #v1 is complete and live; the #mcp2 / 1.0.0 stable-release milestone owns the remaining rollout qualification and soak. The .tasks/ board remains the source of truth.

• 1.1.1 is the current stable release at health.emmetts.dev on Vercel Node 24 + Fluid in iad1. The implementation identity reports 1.1.1 while the MCP protocol SDK remains independently pinned to 2.0.0. The existing SEP-973 icon metadata references the canonical hosted PNG; clients decide whether to render it. The DPoP-capable legacy-auth recovery deployment is retained separately so connector reconnect, Google reconsent, and soak evidence are not conflated with deployment success.

• The MCP endpoint exposes 19 tools and 6 resources through request-scoped SDK v2 transport, with a stateless 2025 fallback for older connectors.

• Google Health consent, encrypted token storage/refresh, identity mapping, reads, writes, audit logging, and freshness behavior have been verified against real data.

• The stable provider's production metadata, RS256 JWKS, public DCR, exact resource/form boundary, scoped MCP challenges, and no-store behavior pass. Existing clients must reconnect once for the new /api/auth/oauth2/* endpoints/scopes; signed-in connector and time-based soak gates remain explicit on #q2 rather than being inferred from anonymous tests.

• Google Health webhook implementation exists under v1.1 (#w11), but its unrelated real Fitbit delivery gate remains open.

Repository structure

.
├── docs/PLAN.md                    ← SOURCE OF TRUTH (architecture, decisions, phases, E2E bar)
├── docs/adr/                       ← accepted architecture decision records
├── README.md                       ← this file (human front door)
├── CLAUDE.md                       ← agent instructions for Claude sessions
├── AGENTS.md                       ← agent instructions, tool-neutral (Codex/Cursor/etc.)
├── .env.example                    ← every env var, documented by name
├── package.json                    ← scripts + pinned deps
├── next.config.ts / tsconfig.json / drizzle.config.ts / vitest.config.ts
│
├── app/                            ← Next.js App Router
│   ├── layout.tsx · page.tsx       ← landing / status dashboard
│   ├── icon.svg                     ← local Google Health MCP favicon
│   ├── sign-in/page.tsx            ← Google sign-in (resumes interrupted OAuth flows)
│   ├── consent/page.tsx            ← private OAuth read/write consent
│   ├── components/sign-out-button.tsx
│   ├── api/
│   │   ├── [transport]/route.ts              ← live MCP endpoint (`/api/mcp`)
│   │   ├── auth/[...all]/route.ts           ← better-auth sign-in + /oauth2 authorize/token/register/userinfo/JWKS
│   │   ├── auth/google-health/start/route.ts    ← health-consent redirect (session-gated)
│   │   ├── auth/google-health/callback/route.ts ← code exchange → encrypt+store → identity map
│   │   └── health/status/route.ts          ← healthcheck (no secrets)
│   └── .well-known/
│       ├── oauth-authorization-server/route.ts
│       └── oauth-protected-resource/route.ts
│
├── src/
│   ├── auth/           ← auth.ts (better-auth config), allowlist, app-user resolution,
│   │                     state (health-consent CSRF state), token-service (single-flight
│   │                     refresh), JWT bearer boundary, DPoP, token-store, google-health-oauth
│   ├── db/             ← domain + legacy/new additive auth schemas, client.ts
│   ├── security/       ← encryption.ts (AES-256-GCM), redact.ts
│   ├── audit/          ← mutation-audit.ts (insert-only audit writer)
│   ├── google-health/  ← typed API client, data-type registry, scopes, and errors
│   ├── time/           ← timezone-safe physical/civil range helpers
│   ├── health-services/ ← shared orchestration layer (also enables future #api)
│   └── mcp/            ← thin tool and resource registration
│
├── drizzle/            ← generated SQL migrations + meta
├── scripts/db-inspect.mjs   ← utility: list/inspect Neon tables
├── tests/unit/         ← default Vitest security, protocol, service, and OAuth coverage
├── tests/integration/  ← opt-in isolated-Postgres atomic credential-replacement proof
└── .tasks/             ← the task board (see "Task board")

The 0.3.0 migration adds isolated mcp_oauth_*_v2 provider tables and google_health_dpop_key. Legacy OAuth client/token/consent tables remain untouched for the seven-day rollback window; health connections, encrypted tokens, caches, webhook data, and audit history are not migrated or deleted.

Setup guide

1. Accounts and services

1. Google Cloud — enable the Google Health API; configure the OAuth consent screen with the googlehealth.* scopes; create an OAuth web client; add the redirect URIs below; publish the app to "In production". Publishing is essential: in Testing status Google expires refresh tokens after 7 days (refresh_token_expires_in: 604799). The "unverified app" badge that appears after publishing is expected and stays — do not submit for verification. Public verification and CASA were evaluated and rejected in ADR-0001.

2. Vercel — import this repo as a project. Framework auto-detects as Next.js. Keep Deployment Protection = preview-only (production must stay open so programmatic MCP clients can reach it; the app enforces its own auth).

3. Neon — the Postgres database was provisioned and connected from the Vercel dashboard through its Storage/Marketplace integration; it is not a separately operated application service. The integration injects DATABASE_URL (pooled) and DATABASE_URL_UNPOOLED automatically. Neon Auth is disabled.

2. Google OAuth redirect URIs

One web client serves both the better-auth sign-in flow (/api/auth/callback/google) and the Google Health consent flow (/api/auth/google-health/callback). Canonical domain: health.emmetts.dev; the Vercel-assigned google-health-mcp-realemmetts.vercel.app also serves. Register all six:

https://health.emmetts.dev/api/auth/callback/google
https://health.emmetts.dev/api/auth/google-health/callback
https://google-health-mcp-realemmetts.vercel.app/api/auth/callback/google
https://google-health-mcp-realemmetts.vercel.app/api/auth/google-health/callback
http://localhost:3000/api/auth/callback/google
http://localhost:3000/api/auth/google-health/callback

Requested health scopes (nine, minimum-necessary — location/ecg/irn are configured on the consent screen but not requested; there is no settings.writeonly):

Read : googlehealth.activity_and_fitness.readonly
       googlehealth.health_metrics_and_measurements.readonly
       googlehealth.sleep.readonly
       googlehealth.nutrition.readonly
       googlehealth.profile.readonly
       googlehealth.settings.readonly
Write: googlehealth.nutrition.writeonly
       googlehealth.health_metrics_and_measurements.writeonly
       googlehealth.profile.writeonly

3. Environment variables

Every variable is documented by name in .env.example. Values are never committed — locally they live in .env.development.local (gitignored); in production they live in the Vercel env store. Never put secret values in code, logs, the task board, or any markdown.

Generate a key: node -e "console.log(require('crypto').randomBytes(32).toString('base64'))"

4. Local development

npm install
cp .env.example .env.development.local     # then fill in values
npm run dev                                # http://localhost:3000
npm test                                   # Vitest
npm run typecheck                          # tsc --noEmit
npm run db:generate                        # generate a Drizzle migration from schema changes
npm run db:migrate                         # apply migrations (uses DATABASE_URL_UNPOOLED)
node scripts/db-inspect.mjs                # list/inspect Neon tables

Local sign-in needs GOOGLE_CLIENT_ID / GOOGLE_CLIENT_SECRET populated and the localhost:3000 redirect URIs registered (both are in place). The dev box is Windows — all npm scripts must be cross-platform.

5. Deploy

Push to main; Vercel builds and deploys. Migrations are run from a dev machine against the unpooled Neon URL (npm run db:migrate), not in the build. Production env-var changes only take runtime effect on the next deploy. Note: an auto-deploy webhook has occasionally not fired on push — if a deploy doesn't appear, deploy manually via the Vercel CLI.

For 0.3.0 sequencing, rollback epochs, fresh-approval boundaries, and cleanup rules, use the 0.3.0 cutover runbook.

Connecting an MCP client

The production endpoint is live. Each client walks the OAuth flow and an approved person signs in with their allowlisted Google account.

Clients previously authorized against /api/auth/mcp/* must remove/re-add the connector once so discovery can use /api/auth/oauth2/*. The public MCP URL itself does not change.

• Claude Code: claude mcp add --transport http health https://health.emmetts.dev/api/mcp

• claude.ai (web + mobile): Settings → Connectors → Add custom connector → https://health.emmetts.dev/api/mcp → complete OAuth.

• ChatGPT: Add a custom connector — it registers itself via Dynamic Client Registration.

MCP surface

Defined in docs/PLAN.md §"MCP surface (v1)"; input schemas per the handoff spec §11. Every Health-data read response carries freshness + units and is payload-bounded (default pageSize ≤ 100; HR series summarized via rollups; truncation notes when capped).

Read/diagnostic tools (13): ping, get_connection_info, get_today_steps, get_sleep_summary, get_latest_heart_rate, get_exercise_week, get_nutrition_log, get_health_context (bundle: sleep + latest HR + resting HR/HRV + recent activity + nutrition — data only, no conclusions), get_health_trends (bounded 7/30/90-day coverage-aware summaries), get_health_updates (durable local notification inbox), query_health_data (generic list/reconcile, registry-allowlisted), rollup_health_data, and get_sync_status.

ping returns the server release, negotiated MCP revision, and auth type. The dedicated get_connection_info tool provides privacy-safe connection diagnostics: client identity and capabilities, all supported protocol revisions, Streamable HTTP/session behavior, OAuth 2.1/DCR/ PKCE/JWT/resource/refresh posture, current scopes/expiry, separate Google Health authorization state, and deployment runtime markers—never credential values, headers, codes, redirects, or email.

Notable behaviors: query_health_data auto-builds the right filter per record type — including the civil date field for daily-* aggregates, which carry no physical timestamp (a sample-time filter would silently return everything); use mode: "reconcile" for Google's merged/deduped stream when multiple sources log the same metric. get_sleep_summary distinguishes STAGES (deep/light/REM) from CLASSIC sessions (a single asleep block — a device capture condition, surfaced via stagesStatus, e.g. REJECTED_COVERAGE, not poor sleep). Staleness is cadence-aware: once-per-night metrics (sleep, daily-*) don't flag isPossiblyStale for a same-day value; live-ish samples use a 3-hour threshold.

Mutation/action tools (6): create_nutrition_log, update_nutrition_log (replace semantics — the live PATCH endpoint 500s; a new data-point name is returned), delete_nutrition_log, create_hydration_log, update_measurement (weight | body-fat | height), and acknowledge_health_updates (local inbox state only). update_profile was dropped: the live endpoint 403s despite the granted scope (documented server-side bug); the service layer is kept for re-enablement.

Resources (6): health://profile, health://settings, health://connected-user, health://data-types, health://freshness, health://updates.

Absent by design: sleep, exercise, and settings writes; bulk historical writes.

Security posture

• Four auth layers, never conflated (table above). The MCP endpoint requires a valid exact-audience JWT with health:read (and health:write before mutations); signature, issuer, audience, expiry, subject, scope, and current email allowlist are checked locally. Sign-in is allowlisted to ALLOWED_GOOGLE_EMAILS; DCR is open by design (any client may register — safety comes from the allowlisted login, not registration). The approved identity and the prohibition on public signup are fixed by ADR-0002.

• Google tokens and DPoP private keys are AES-256-GCM encrypted at rest (TOKEN_ENCRYPTION_KEY, with purpose-separated HKDF context and key_version for rotation). No plaintext tokens in the DB, logs, or error paths — a redact() helper strips token patterns (ya29., 1//, GOCSPX-, JWTs, Bearer/Basic, Neon npg_) before anything is logged.

• MCP credentials use a different storage model: provider refresh tokens and public-client secrets are hashed; MCP access JWT values are not persisted. A local JWKS cache verifies warm requests without a token-table lookup.

• Token refresh is single-flight through an atomic, expiring database claim on the token row (compatible with the stateless Neon HTTP driver); on refresh failure the connection is marked reauth_required.

• The allowlist is rechecked on every MCP bearer request. Removal blocks tool use immediately; complete offboarding still deletes Better Auth sessions/MCP grants and the person's stored Google Health connection.

• Writes (nutrition / hydration / measurements only) are Zod-validated, explicit-input only, and every mutation is audit-logged in mutation_audit_log. There are no sleep / exercise / settings write tools at all.

• No medical diagnosis language anywhere; freshness/limitation notes on every response.

• Production stays open (no Vercel SSO wall) on purpose — the application's own auth is the perimeter.

• Stable OAuth Provider 1.6.25 has three tracked residuals: its resource-indicator advisory is contained by one configured audience plus exact resource/audience boundaries. A refresh-only compatibility boundary inserts the canonical resource when a client omits it, but only while that canonical endpoint is the complete configured resource set; supplied invalid values still fail closed. Refresh rotation uses predecessor compare-and-set followed by successor insertion rather than one provider transaction. Local/preview auth storage also shares the Vercel-linked Neon database; preview is Vercel-protected and only the allowlisted owner can grant. Re-evaluate all three when a stable provider fix or environment-isolated auth store is adopted.

Troubleshooting

Claude work-computer OAuth: browser returns, connector still fails

This is the MCP client-auth flow, not Google Health consent. If the browser already has a valid Google session, clicking Sign in with Google can complete without showing Google's account chooser. The saved MCP authorization request then resumes and returns to Claude immediately. That silent SSO path is normal; the useful question is what happened after the return.

Anthropic's remote-connector guide explains that remote connectors are configured and brokered through the Claude account, with the MCP connection originating from Anthropic's servers rather than the workstation. They can be shared across Claude on the web, Desktop, and Claude Code when the same account and workspace are selected; see Anthropic's connector guidance. On Team and Enterprise plans, an Owner or Primary Owner must add/enable the connector for the organization, and each user normally authenticates it individually.

Record the last confirmed stage before changing anything:

For the work-computer incident that prompted this runbook, production evidence showed Claude access tokens being persisted after the browser returned. A regression replay then found that better-auth's deprecated MCP path issued an ephemeral HS256 ID token while advertising RS256, omitted the issuer, wrote auth_time in milliseconds, and advertised JWKS/UserInfo endpoints that returned 404. A strict client can reject that response before its first Bearer request even though the server has already stored the access token. Version 0.1.2 repairs that boundary with an encrypted, persisted RS256 signing key and live JWKS/UserInfo endpoints; the work-computer retry remains the final acceptance check. This is MCP client auth, not evidence that Google Health consent failed.

That paragraph describes the pre-0.3.0 legacy bridge retained only as rollback data. New connections use the maintained OAuth Provider, persisted RS256 JWKS, exact /api/mcp audience, and locally verified JWT bearer requests.

A fresh 2026-07-14 revalidation also completed native DCR, S256 PKCE, browser handoff, loopback callback, token exchange, MCP initialize/discovery, and a read-only tool call through both Claude Code and Codex against production. Claude used a localhost callback; Codex used its 127.0.0.1:<ephemeral>/callback/<server-id> shape. The account-level Claude connector and the separately authenticated direct Claude Code entry were both connected, and the correlated server window contained no relevant OAuth/MCP 5xx or unexpected 4xx responses. A new failure that occurs before any server request is therefore on the client, work-network, or workspace-policy side, not an MCP callback failure demonstrated at the Health server.

Do not rotate shared auth/encryption secrets, revoke unrelated users' or clients' tokens, disconnect Google Health, or clear every credential as a first response. Scope any reset to the one affected connector.

Recovery sequence:

1. Capture the exact Claude error (including any ofid_... identifier), local timestamp with timezone, Claude Code or Desktop version, selected Claude account/workspace, and whether the connector came from Claude's account settings or a direct Claude Code MCP entry.

2. Update Claude Code/Desktop, fully quit the affected clients (including the Desktop tray process), reopen them, and confirm the same account and workspace are selected. Use the latest available release; see Anthropic's Claude Code changelog.

3. Check the account-level connector in Claude's connector settings. On Team/Enterprise, have an Owner or Primary Owner confirm its organization configuration. Disconnect/reconnect the affected user, or remove and re-add that remote connector once; avoid repeated retries that create more client registrations without isolating the failing stage. Re-add it only at https://health.emmetts.dev/api/mcp, authenticate once, and test it on Claude web before checking Desktop and Code.

4. In Claude Code, check claude mcp list and /mcp for a separately configured direct entry pointing to the same URL. Direct local/project/user entries take precedence over a duplicate claude.ai connector. For that duplicate direct entry only, use Clear authentication in /mcp, remove the duplicate, and retry with the account connector (or re-add the direct entry once if it is intentionally preferred). Anthropic documents this precedence and the clear-auth control in the Claude Code MCP guide.

5. Use Claude Code's “paste the full callback URL” fallback only when the browser ends on a literal connection error at a local http://localhost:<port>/callback URL and Claude Code is prompting for that URL. It does not apply to a normal remote return to Claude Desktop or to a callback that already completed.

6. If the failure persists on 0.1.2 after the ID token verifies against the advertised JWKS, preserve the captured ofid_..., timestamp, versions, workspace type, and sanitized /mcp status for Anthropic support, together with the server-side timeline. A valid token response followed by no authenticated MCP request is then a post-token client/connector failure.

For Codex, codex mcp list should show the Health entry as OAuth; use codex mcp login shaughv-health for a connector-scoped retry. Codex registers a public client and owns a randomized 127.0.0.1 callback, so no fixed Codex callback belongs in Google Cloud. If a Codex run fails before the server receives DCR or an unauthenticated MCP discovery probe, update Codex and correct the local model/client configuration first. See OpenAI's Codex MCP guide.

Hermes Desktop, direct Claude Code, and Codex use different local callback shapes, all supported by this server: Hermes uses an ephemeral 127.0.0.1:<port>/callback, Claude Code normally uses its fixed localhost callback port, and Codex uses an ephemeral 127.0.0.1 path that can include a server-specific callback identifier. Native IP-literal loopback redirects accept the actual listening port selected at authorization time; the protocol, host, path, and query remain exact.

The protected-resource document and initial 401 advertise the complete approved initial grant: openid profile email offline_access health:read health:write. Current clients differ in whether they treat the challenge, protected-resource document, or authorization-server metadata as the authoritative scope set, so all three now agree. This intentionally includes offline_access in resource metadata for uninterrupted private-client refresh interoperability. A real write-scope failure still returns only the explicit health:write step-up challenge. Existing public registrations are expanded additively, but an already issued one-hour access token does not become refreshable: reauthenticate that connector once to receive a rotating refresh credential.

Task board

Work is tracked on a self-contained board under .tasks/ (the SHAUGHV tasks system).

• .tasks/TASKS.md — the board (Backlog / To-Do / Active / Done). Source of truth for what's next.

• .tasks/MILESTONES.md — dated epics; tasks join one with an (ms #id) tag. Current milestone: #mcp2 (1.0.0 stable release and remaining qualification).

• .tasks/tasks/<id>.md — a rich detail file per task (TT;DR-led, with ## Verification, ## Status, ## Activity). The decision history lives here — e.g. rlw.md (Railway/FastMCP), api.md (REST surface feasibility), inf.md (the infra session), w11.md (webhooks v1.1).

• .tasks/CLAUDE.md — working memory (people, terms, projects, preferences).

The live dashboard is a zero-dependency Node server; resolve its port from .tasks/.board-server.json or run node .tasks/board-server.mjs status. See CLAUDE.md / AGENTS.md for how agents should use the board.

Further reading

• docs/adr/0001-private-allowlist-only.md — accepted private-audience, OAuth verification, and access-control decision.

• docs/PLAN.md — the authoritative build plan (read first).

• docs/adr/0003-vercel-node-fluid-mcp-2026.md — current host/runtime decision.

• docs/research/2026-07-29-mcp-sdk-v2-and-hosting.md — SDK v2, auth, hosting, compatibility, and pricing research.

• CLAUDE.md — instructions for Claude coding sessions.

• AGENTS.md — the same operational guidance, tool-neutral, for any agent (Codex, Cursor, etc.).

• .env.example — the env-var reference.