feat(skills): add mcp-oauth-remote-gateway skill

Add an optional skill for connecting OAuth-gated remote MCP servers (Better Stack, Linear, Cloudflare, Datadog, Stripe, etc.) when Hermes runs as a remote gateway, where the built-in browser OAuth flow cannot capture the 127.0.0.1 callback. Covers the manual RFC 7591 DCR + RFC 7636 PKCE + authorization_code flow, writing tokens in Hermes' HermesTokenStorage schema, the dashboard-first escalation path, and a diagnostic script + pitfalls for refresh/session-revocation recovery.
2026-07-19 15:18:03 +00:00 · 2026-07-08 07:57:18 +00:00 · 2026-07-08 07:57:18 +00:00 · 03885e0aa1
commit 03885e0aa1
parent 49a8c3f836
3 changed files with 588 additions and 0 deletions
--- a/skills/mcp/mcp-oauth-remote-gateway/SKILL.md
+++ b/skills/mcp/mcp-oauth-remote-gateway/SKILL.md
@ -0,0 +1,348 @@
+---
+name: mcp-oauth-remote-gateway
+description: Connect an OAuth-gated remote MCP server (Better Stack, Linear, Cloudflare, Datadog, Stripe, etc.) to Hermes when Hermes runs as a REMOTE gateway (container, VPS, messaging bot) where the built-in browser OAuth flow cannot capture the localhost callback. Manually performs RFC 7591 Dynamic Client Registration + RFC 7636 PKCE + authorization_code exchange and writes tokens in the exact schema Hermes persistence expects.
+version: 1.0.0
+author: Hermes Agent
+license: MIT
+metadata:
+  hermes:
+    tags: [MCP, OAuth, PKCE, remote-deployment]
+    related_skills: [native-mcp, mcporter]
+---
+
+# MCP OAuth on a Remote Hermes Gateway
+
+## Overview
+
+Hermes' built-in MCP OAuth client runs a one-shot HTTP listener on `127.0.0.1:<port>`
+inside the Hermes process and registers that loopback address as the OAuth
+`redirect_uri`. That works perfectly for a local CLI on the user's own machine.
+It breaks completely when Hermes runs as a remote gateway (container, VPS,
+messaging bot), because the user's browser resolves `127.0.0.1` to the user's own
+laptop, not the remote container — so the authorization code never reaches Hermes.
+
+This skill does the OAuth dance by hand and writes the resulting tokens into the
+exact files Hermes' token storage expects, so a subsequent `/reload-mcp` finds
+cached tokens and skips the browser flow entirely.
+
+## When to Use
+
+Use this skill when **all** of the following are true:
+
+1. The user wants to add a remote HTTP MCP server that requires OAuth (not a static Bearer token).
+2. Hermes is running as a **remote gateway** (container, VPS, Docker, managed service) — NOT a local CLI on the user's laptop.
+3. The server supports OAuth 2.1 with PKCE and RFC 7591 Dynamic Client Registration (most modern MCP servers do — Better Stack, Linear, Cloudflare, Datadog, etc.). If it doesn't support DCR (GitHub is the notable exception), this skill does not apply — use a pre-registered OAuth App or a Personal Access Token instead.
+
+Do NOT use this for:
+- **Local CLI Hermes** — just set `auth: oauth` in `mcp_servers.<name>` and `/reload-mcp`. The built-in flow opens a browser and captures the callback on localhost. Works perfectly.
+- **Servers that accept a static Bearer token (API key)** — always prefer `headers.Authorization: "Bearer <token>"` when the user is willing. Simpler, no refresh dance.
+- **GitHub Copilot MCP** (`api.githubcopilot.com/mcp/`) — GitHub does not expose DCR. Use a PAT or a pre-registered OAuth App (see pitfall 12).
+
+## Why the Built-in OAuth Flow Fails on a Remote Gateway
+
+Hermes' native MCP OAuth client (`tools/mcp_oauth.py`):
+
+1. Picks a free local port `P`.
+2. Registers a dynamic OAuth client with the AS, sending `redirect_uri = http://127.0.0.1:P/callback`.
+3. Starts an HTTP server on `127.0.0.1:P` **inside the Hermes process**.
+4. Prints the authorize URL and waits for the code at its local endpoint.
+
+When Hermes runs remotely, the `127.0.0.1` in the `redirect_uri` is the remote
+container's loopback, not the user's. After authorizing, the user's browser 302s
+to `http://127.0.0.1:P/callback?code=...`, which resolves to the user's own
+laptop and fails to connect. The callback never reaches the Hermes process, the
+flow times out, and `/reload-mcp` returns "No MCP tools available" with no detail.
+
+Symptoms to recognize: `[xdg-open] <defunct>` processes under the hermes user, an
+empty or missing tokens directory (`$HERMES_HOME/mcp-tokens/`), and a reload that
+responds without any "Added/Reconnected: X" line in `change_detail`.
+
+## Preferred Front Door: the Hermes Dashboard (try this BEFORE manual token surgery)
+
+A remote Hermes gateway often also runs the **dashboard** web UI as a SEPARATE
+process (e.g. `hermes dashboard --host 0.0.0.0 --port <port>`; check with
+`ps aux | grep 'hermes dashboard'`). It exposes a connector/MCP console —
+endpoints like `/api/mcp/servers`, `/api/mcp/status`, and `/connectors` (all
+login-gated; a cookieless curl returning 401/302 confirms they exist).
+
+**Why the dashboard solves the core problem:** when the user drives OAuth from
+the dashboard *in their own browser*, the redirect lands in a context the
+dashboard can capture — sidestepping the `127.0.0.1`-callback failure that breaks
+the CLI/manual flow. So the correct escalation order for "add or re-auth an OAuth
+MCP server on a remote gateway" is:
+
+1. **Dashboard, in the user's browser** — the intended front door. Add servers, run OAuth, reload, all authenticated as the user. No copy-paste-callback dance, no hand-writing token files.
+2. **Manual token surgery (the rest of this skill)** — the FALLBACK for when there's no browser session to the dashboard (pure-chat/headless context).
+
+**Finding the dashboard's PUBLIC URL.** The dashboard binds internally to
+`0.0.0.0:<port>`, but the user needs the externally-reachable URL. Most deploy
+platforms inject it into the environment — grep for it rather than making the
+user hunt:
+
+```bash
+env | grep -iE "HERMES_DASHBOARD_PUBLIC_URL|RAILWAY_PUBLIC_DOMAIN|RAILWAY_STATIC_URL|RAILWAY_SERVICE_.*_URL|PUBLIC_URL|BASE_URL|DOMAIN" \
+  | sed -E 's/(TOKEN|SECRET|KEY|PASSWORD)=.*/\1=***REDACTED***/I'
+```
+
+`HERMES_DASHBOARD_PUBLIC_URL` is authoritative when present. On Railway also check
+`RAILWAY_PUBLIC_DOMAIN` / `RAILWAY_STATIC_URL` (the `*.up.railway.app` host) and
+`RAILWAY_SERVICE_*_URL` vars, which sometimes carry a friendlier custom domain.
+Hand the user the full `https://` URL and point them at the Connectors/MCP
+section. ALWAYS pipe through the `sed` redaction above — these env greps sit next
+to `*_TOKEN`/`*_SECRET` vars.
+
+**What the dashboard does NOT fix (still host-side / shell):** stdio servers that
+need shell auth state (a CLI `login` command whose credentials may not persist
+across restarts) and anything reading credentials from `$HERMES_HOME/.env`. Those
+are out of the dashboard's scope regardless.
+
+## The Workaround
+
+Do the OAuth dance manually, then write the resulting tokens into the exact files
+Hermes' `HermesTokenStorage` would have written, so on `/reload-mcp` Hermes finds
+cached tokens and skips the browser flow entirely.
+
+### 1. Confirm it's a remote gateway
+
+```bash
+env | grep -iE "HERMES|RAILWAY|CONTAINER"
+echo "$DISPLAY $WAYLAND_DISPLAY $SSH_CLIENT"
+```
+
+No display + a remote indicator = remote gateway. `tools/mcp_oauth.py::_can_open_browser()`
+uses these same env vars, so if Hermes' own auto-detect says "headless", the
+built-in flow won't work.
+
+### 2. Find HERMES_HOME and the config path
+
+```bash
+HERMES_HOME=$(python3 -c 'from hermes_constants import get_hermes_home; print(get_hermes_home())')
+echo "config: $HERMES_HOME/config.yaml"
+echo "tokens: $HERMES_HOME/mcp-tokens/"
+```
+
+### 3. Discover OAuth metadata from the MCP server
+
+MCP servers advertise their OAuth setup via RFC 9728 (OAuth 2.0 Protected
+Resource Metadata). The `WWW-Authenticate` header on a 401 tells you where to look:
+
+```bash
+curl -sI https://mcp.example.com | grep -i www-authenticate
+# → Bearer realm="mcp", resource_metadata="https://mcp.example.com/.well-known/oauth-protected-resource"
+```
+
+**Not every server returns `WWW-Authenticate`.** Some return a bare
+`{"errors":["Unauthorized"]}` 401 with no auth-discovery hint. When that happens,
+probe well-known paths directly:
+
+```bash
+for p in \
+  /.well-known/oauth-protected-resource \
+  /.well-known/oauth-authorization-server \
+  /.well-known/openid-configuration ; do
+  echo "=== $p ==="
+  curl -s -A "python-httpx/0.27" "https://mcp.example.com$p" | head -c 400; echo
+done
+```
+
+Fetch the resource metadata to get `authorization_servers`, then fetch the AS's
+`/.well-known/oauth-authorization-server` to get `authorization_endpoint`,
+`token_endpoint`, and `registration_endpoint`.
+
+Pitfall: many servers sit behind Cloudflare and 403 bare `urllib` user agents.
+Always set `User-Agent: python-httpx/0.27` (or similar) on requests in this flow.
+
+### 4. Dynamic Client Registration (RFC 7591)
+
+POST to the `registration_endpoint` with:
+
+```json
+{
+  "client_name": "Hermes Agent (manual OAuth)",
+  "redirect_uris": ["http://127.0.0.1:8765/callback"],
+  "grant_types": ["authorization_code", "refresh_token"],
+  "response_types": ["code"],
+  "token_endpoint_auth_method": "none",
+  "scope": "<scopes_from_resource_metadata>"
+}
+```
+
+Omit `scope` entirely if the AS's `scopes_supported` is empty — see step 5
+pitfall. Use port `8765` (or any port — nothing will listen).
+`token_endpoint_auth_method: none` marks this as a public PKCE client. Save the
+returned `client_id`.
+
+### 5. Build the authorize URL with PKCE
+
+Generate:
+- `code_verifier`: `secrets.token_urlsafe(64)[:128]`
+- `code_challenge`: `base64url(sha256(code_verifier))` (no padding)
+- `state`: `secrets.token_urlsafe(24)`
+
+Query params: `response_type=code`, `client_id`, `redirect_uri`, `code_challenge`,
+`code_challenge_method=S256`, `state`, plus `resource=<mcp_server_url>` (RFC 8707 —
+many servers require this to bind the token to the specific MCP resource). Include
+`scope=<space-separated>` ONLY if the AS metadata's `scopes_supported` is a
+non-empty array AND/OR the resource metadata declares specific scopes. If
+`scopes_supported: []`, omit the `scope` parameter — the server grants its full
+default set on its own. Fabricating scope strings against an empty
+`scopes_supported` can cause `invalid_scope` errors on some ASes.
+
+**Stash `code_verifier` and `state` to disk** (e.g. `/tmp/.mcp-oauth-work/<server>.json`,
+0600 perms). You need them for step 7, possibly across multiple chat turns.
+
+### 6. Give the user the authorize URL
+
+```
+Open this URL in your browser:
+<authorize_url>
+
+After approving, your browser will try to load http://127.0.0.1:8765/callback
+and fail to connect — THAT'S EXPECTED. Just copy the entire URL from the
+address bar (it will contain ?code=...&state=...) and paste it back here.
+```
+
+### 7. Exchange the code for tokens
+
+When the user pastes the callback URL:
+
+1. Parse `code` and `state` from the query string.
+2. **Verify `state` matches the stashed value** (CSRF check — do not skip).
+3. POST `application/x-www-form-urlencoded` to the `token_endpoint`:
+   - `grant_type=authorization_code`
+   - `code=<from callback>`
+   - `redirect_uri=<same as step 4>`
+   - `client_id=<from step 4>`
+   - `code_verifier=<stashed>`
+   - `resource=<mcp_server_url>` (if the AS required it in step 5, include here too)
+4. Response contains `access_token`, `refresh_token`, `token_type`, `expires_in`, `scope`.
+
+### 8. Write tokens in Hermes' exact schema
+
+`tools/mcp_oauth.py::HermesTokenStorage` expects two files under
+`$HERMES_HOME/mcp-tokens/` (create dir with `0o700`, files with `0o600`):
+
+**`<server_name>.json`** — the `OAuthToken` pydantic model:
+```json
+{
+  "access_token": "...",
+  "token_type": "Bearer",
+  "expires_in": 7200,
+  "refresh_token": "...",
+  "scope": "read write"
+}
+```
+
+**`<server_name>.client.json`** — the `OAuthClientInformationFull` model:
+```json
+{
+  "client_id": "...",
+  "redirect_uris": ["http://127.0.0.1:8765/callback"],
+  "grant_types": ["authorization_code", "refresh_token"],
+  "response_types": ["code"],
+  "token_endpoint_auth_method": "none",
+  "scope": "read write",
+  "client_name": "..."
+}
+```
+
+Write each file via `json.dumps(..., indent=2)`. Sanitize the filename with
+`re.sub(r'[^\w\-]', '_', server_name)[:128]` — this matches `_safe_filename()` in
+Hermes' token storage.
+
+### 9. Add the server to config.yaml
+
+```yaml
+mcp_servers:
+  <name>:
+    url: "https://mcp.example.com"
+    auth: oauth
+    timeout: 180
+    connect_timeout: 60
+```
+
+### 10. Smoke-test the token BEFORE asking the user to reload
+
+Manually POST an MCP `initialize` request to confirm the token works end-to-end —
+this catches scope misconfigurations, wrong `resource` values, and CF blocks
+before the user is confused by another "No MCP tools available" reload:
+
+```python
+body = json.dumps({
+    "jsonrpc": "2.0", "id": 1, "method": "initialize",
+    "params": {
+        "protocolVersion": "2025-06-18",
+        "capabilities": {},
+        "clientInfo": {"name": "hermes-debug", "version": "1.0"},
+    },
+}).encode()
+# POST to the MCP URL with:
+#   Authorization: Bearer <access_token>
+#   Accept: application/json, text/event-stream
+#   Content-Type: application/json
+#   MCP-Protocol-Version: 2025-06-18
+#   User-Agent: python-httpx/0.27
+```
+
+Expect HTTP 200 with `Content-Type: text/event-stream` and a JSON-RPC result
+containing `serverInfo` and `capabilities`. **Do not use `urllib` with its default
+UA** — Cloudflare will 403 you even though Hermes (which uses httpx) will succeed.
+`scripts/diagnose-oauth-mcp.py` automates this smoke test.
+
+### 11. Tell the user to run `/reload-mcp`
+
+On reload, Hermes sees `auth: oauth`, calls `HermesTokenStorage.get_tokens()`,
+finds your cached tokens, skips the browser flow, and registers `mcp_<name>_*`
+tools. Refresh happens automatically before `expires_in` elapses.
+
+## Pitfalls & Lessons Learned
+
+1. **Do not assume "headless" means "OAuth impossible."** The built-in flow works fine for local CLI; the issue is strictly remote deployments where the user's browser and the Hermes process are on different machines. Check the execution environment before claiming OAuth isn't an option.
+
+2. **Read the source, not just the skill docs.** `tools/mcp_oauth.py` and the MCP config reference in `website/docs/` are the authoritative references. Grep the tree before telling the user a feature "doesn't exist."
+
+3. **Cloudflare UA filter.** Many MCP/OAuth providers front their infra with Cloudflare, which 403s `python-urllib/*` user agents on metadata endpoints even though those endpoints are public. Set `User-Agent: python-httpx/0.27` (or any browser-like string) on every request in this flow. Hermes itself uses httpx, so this is never a problem in the real connection path.
+
+4. **Include `resource` in both authorize and token requests.** RFC 8707 resource indicators are not optional for most modern MCP servers — they bind the issued token to the specific MCP resource URL. Leaving it out sometimes still works but may yield a token that later fails at the MCP server with a scope/audience error.
+
+5. **Trailing slash matters.** Some servers advertise the resource as `https://mcp.example.com/` with a trailing slash and reject tokens issued against the no-slash variant. Copy the `resource` value verbatim from the `.well-known/oauth-protected-resource` response.
+
+6. **`/reload-mcp` is silent on failure.** If the reload shows "No MCP tools available" with no `change_detail` line, a server is in config but failed to connect and no error bubbled up. Tail the error log, smoke-test the token directly with a manual `initialize` POST, and — if everything looks good — ask for a full process restart.
+
+7. **Circuit breaker can survive `/reload-mcp`.** `tools/mcp_tool.py` keeps a module-level error-count dict with a small threshold. Once tripped (e.g. after token expiry produces several consecutive failures), the tool handler can short-circuit before calling the server, so no successful call resets the counter. Symptom: reload says "Reconnected: X" but subsequent calls still fail with "server unreachable" in the same conversation. Recovery order: try `/reload-mcp` FIRST (cheap, no chat-process blip) — on current builds it can clear the counter; only escalate to a full gateway process restart if a live call STILL short-circuits after reload. Do not lead with "you must restart."
+
+8. **Refresh on an expired access_token + a tripped breaker is a deadlock.** The auto-refresh logic runs inside the MCP call path, which the breaker short-circuits once tripped. Manually refreshing the token on disk does not help by itself — pair a manual token refresh with a full restart, not a `/reload-mcp`.
+
+9. **`invalid_grant` on a manual refresh means the refresh token is DEAD — re-auth is the only fix, do not loop.** When the access_token has been expired long enough, the refresh_token can also be revoked/expired server-side. A `grant_type=refresh_token` POST then returns HTTP 400 `{"error":"invalid_grant",...}` (wording varies: "Grant not found", "Token expired", "refresh token is invalid"). There is NO recovery from the gateway side. Hand back to the user with two options: (a) re-run the full manual OAuth dance (steps 3–10), or (b) if the provider offers a static personal API key, switch to that — no refresh/expiry cycle, more durable for an unattended remote gateway. Detect early: before any create/update operation against an OAuth MCP, check `expires_at` vs `time.time()`; if already expired, attempt the refresh first and surface `invalid_grant` immediately rather than failing mid-task.
+
+10. **A successful refresh that STILL yields a rejected token = server-side SESSION revocation; only a fresh authorization_code flow fixes it.** Distinct from pitfall 9. The stored token file can look healthy (`expires_at` well out, refresh_token present), yet a live `initialize` POST returns `401 invalid_token` with a JSON-RPC body like `{"error":{"code":-32002,"message":"Session expired. Please re-authenticate."}}`. The `grant_type=refresh_token` POST may **succeed** (HTTP 200, new access_token) — yet the brand-new token gets the SAME `-32002`. The provider revoked the underlying MCP *session* server-side; the OAuth refresh chain re-mints credentials but cannot re-establish a revoked session. Decision rule when an OAuth MCP reports "not connected": (1) smoke-test the stored access_token with a manual `initialize` POST; (2) if `401 invalid_token`, attempt a refresh and smoke-test the NEW token; (3a) new token works → write it + restart to clear the breaker; (3b) new token STILL gets `-32002`/"Session expired" → stop, this is session revocation, hand the user the authorize URL for a full re-auth. `scripts/diagnose-oauth-mcp.py` automates steps 1–2 and prints which branch you're in. For an unattended gateway whose session keeps getting revoked, prefer a static Personal API key. See `references/stripe-mcp-oauth-revocation.md` for a worked example of a provider that revokes weekly.
+
+11. **Client info file is NOT optional.** Hermes needs `<server>.client.json` to know the `client_id` for refresh grants. Skipping it means the first refresh fails and the user has to re-auth — writing both files is the whole point of this skill.
+
+12. **Never hand-type the redirect URL for the user to open.** Generate the authorize URL programmatically with `urllib.parse.urlencode()`. Spaces in scopes and special chars in `state` break string-concatenated URLs.
+
+13. **Security: the stash file contains the `code_verifier`.** Delete `/tmp/.mcp-oauth-work/<server>.json` immediately after successful token exchange. There's no reason to keep a proof-of-identity secret around once it's consumed.
+
+14. **Write what the token endpoint actually returned.** The AS may grant a narrower (or wider) scope than requested. Write the `scope` from the token-exchange response to `<server>.json`, not what you asked for in step 5. When `scopes_supported: []`, the explicit scope list you send IS authoritative both ways: some servers grant exactly what you list (pass narrow scopes for least-privilege, or enumerate the full set if the user needs everything), and some won't echo the granted scope back at registration time — only the token-exchange response is authoritative.
+
+15. **OAuth tokens often double as Bearer tokens against the provider's public REST API.** The access_token in `<server>.json` is frequently not "MCP-only" — `Authorization: Bearer <token>` against the provider's documented REST API succeeds whenever the corresponding resource scope was granted. This is the OAuth 2.0 spec, not a provider quirk. When the MCP server is read-only but you need a write operation, check whether the OAuth token can hit the provider's REST API directly before suggesting a separate API key.
+
+16. **Secret redaction can mask tokens in tool output.** If secret redaction is enabled, tokens and long opaque strings render as `***` in tool-result output, so you cannot `print(response)` to keep the access_token visible across turns. Combined with single-use `code` values from authorization_code grants: if you print the token-exchange response, you may lose the token AND consume the code, forcing a restart with a fresh authorize URL. **Always write the access_token directly to its final destination file in the SAME code block that performs the token exchange.** If you must print for debugging, print only `len(access_token)`, `token_type`, `scope`, `expires_in` — never the secret.
+
+17. **GitHub MCP (`api.githubcopilot.com/mcp/`) uses a pre-registered confidential OAuth App, not DCR + PKCE-public.** Its client info ships with a real `client_secret` and `token_endpoint_auth_method: client_secret_post`. The token-exchange POST to `https://github.com/login/oauth/access_token` must include `client_secret` as a form field alongside `client_id`, `code`, `code_verifier`, and `redirect_uri` (PKCE is still honored on top of the secret). The redirect URI is **fixed** in the OAuth App config — you cannot change it, so the manual listener-port trick doesn't apply; the user just lets the browser fail to connect on that port and pastes the address-bar URL back.
+
+## What NOT to do
+
+- **Don't use `mcp-remote` as a fallback.** It runs an npx subprocess whose OAuth callback server ALSO sits on the remote container's localhost — same problem. `mcp-remote` only helps when the MCP client doesn't speak remote HTTP at all (Hermes does natively).
+- **Don't push "paste your API token and I'll add headers"** if the user explicitly asked for OAuth. Offer the static-token shortcut only after explaining why the native OAuth flow fails in remote deployments. Respect the user's choice to do the extra legwork for rotation-free, scope-limited access.
+- **Don't claim Hermes doesn't support a feature without reading the source.** Grep the source tree before making capability claims.
+
+## Quick Reference Files
+
+- `scripts/diagnose-oauth-mcp.py` — re-runnable, read-only-by-default diagnostic. Given a server name, it smoke-tests the stored access_token, attempts a refresh, smoke-tests the new token, and prints exactly which recovery branch you're in (`TOKEN_OK` = breaker/restart, `REFRESH_FIXED` = persist+restart, `SESSION_REVOKED` = full re-auth, `REFRESH_DEAD` = full re-auth/API key). Pass `--write` to persist a working refreshed token atomically. Never prints secret values. **Run this FIRST when an OAuth MCP server reports "not connected"** — it encodes the pitfall 7/9/10 decision tree.
+- `references/stripe-mcp-oauth-revocation.md` — a worked example (Stripe) of a provider that revokes its OAuth session on a recurring basis, and the durable fix: switch to a static restricted API key.
+
+## Related
+
+- `native-mcp` — general guide to configuring MCP in Hermes. Authoritative config reference lives there.
+- `mcporter` — the external CLI bridge, for ad-hoc MCP calls outside of Hermes' config.
--- a/skills/mcp/mcp-oauth-remote-gateway/references/stripe-mcp-oauth-revocation.md
+++ b/skills/mcp/mcp-oauth-remote-gateway/references/stripe-mcp-oauth-revocation.md
@ -0,0 +1,62 @@
+# Stripe MCP (`mcp.stripe.com`) — recurring OAuth session revocation, fix with a restricted key
+
+A worked example of pitfall 9/10 in SKILL.md: a provider whose OAuth session dies
+on a recurring basis, where the durable fix is to drop OAuth for a static API key.
+
+## Symptom (the "dies after a while" complaint)
+Stripe MCP works for a few days then goes "not connected." Auto-refresh is healthy
+in between (the access token is 1h-lived and rotates fine), so it LOOKS like a
+refresh-token expiry or a max-session cap — it is neither. Roughly weekly, Stripe
+**revokes the entire OAuth grant server-side**. The next `grant_type=refresh_token`
+POST returns:
+
+```
+HTTP 400  {"error":"invalid_grant","error_description":"Invalid refresh token"}
+```
+
+The whole grant is dead — not just the short-lived access token — so auto-refresh
+cannot recover it. It requires a fresh interactive browser consent flow, which a
+headless remote gateway cannot drive. Don't be fooled by a green smoke-test at any
+given moment: the failure is intermittent revocation, not a permanently broken token.
+
+## Why the three usual hypotheses are all wrong
+Per Stripe's OAuth docs (https://docs.stripe.com/stripe-apps/api-authentication/oauth):
+- **Access tokens** expire in **1 hour**.
+- **Refresh tokens** expire after **1 year**, and are **rolled on every exchange** — so
+  as long as you refresh at least once a year they never naturally expire.
+- Hermes auto-refreshes independently of whether you call Stripe tools, so "not using
+  the tools enough" is irrelevant.
+
+So a *recurring* death cannot be refresh-token expiry (1yr) or "max OAuth session
+length" (no clean documented cap). It is server-side session revocation. Do NOT
+loop on refresh.
+
+## The durable fix: drop OAuth, use a restricted API key as a Bearer token
+Stripe's MCP docs (https://docs.stripe.com/mcp) are explicit that for
+non-interactive / agent use, OAuth is the wrong tool — `mcp.stripe.com` accepts a
+**static restricted key** (`rk_live_...`) as a Bearer token. A restricted key has
+**no session, no refresh, no expiry** — it works until revoked, ending the re-auth
+cycle entirely.
+
+config.yaml change (no token files needed — delete the OAuth dance for this server):
+```yaml
+mcp_servers:
+  stripe:
+    url: https://mcp.stripe.com
+    headers:
+      Authorization: *** rk_live_..."      # restricted key from Dashboard
+      # Stripe-Account: "acct_xxx"            # only for Connect platform / connected-account calls
+```
+
+Generate the key in Stripe Dashboard → Developers → API keys → **Restricted keys**.
+Grant least-privilege scopes for what the bot actually does:
+- account reads: **read** on Charges, Customers, Subscriptions, Coupons/Promotion codes
+- refunds / writes: add the corresponding **write** scopes
+Then `/reload-mcp` (full restart only if the breaker is tripped, per pitfall 7).
+
+## Decision rule
+For ANY unattended remote-gateway MCP server that keeps getting its session revoked
+(`invalid_grant` on refresh, or `-32002 "Session expired"` after a successful refresh),
+and whose provider offers a static API key — prefer the static key over OAuth. OAuth's
+refresh dance is for interactive clients; it is a liability for a headless gateway.
+Stripe (restricted key) and Linear (Personal API key) both fit this rule.
--- a/skills/mcp/mcp-oauth-remote-gateway/scripts/diagnose-oauth-mcp.py
+++ b/skills/mcp/mcp-oauth-remote-gateway/scripts/diagnose-oauth-mcp.py
@ -0,0 +1,178 @@
+#!/usr/bin/env python3
+"""Diagnose an OAuth-gated remote MCP server's connection state.
+
+Decides which recovery branch you're in WITHOUT mutating disk by default:
+  1. Smoke-test the stored access_token against the MCP endpoint (initialize).
+  2. If 401, attempt a refresh and smoke-test the freshly-minted token.
+
+Branches printed at the end:
+  TOKEN_OK            -> stored token works; "not connected" is the circuit
+                         breaker (SKILL pitfall 7) -> restart the gateway.
+  REFRESH_FIXED       -> refresh minted a working token; pass --write to persist
+                         it (atomic, 0600), then restart to clear the breaker.
+  SESSION_REVOKED     -> refresh succeeds but new token STILL gets -32002
+                         "Session expired" (pitfall 10) -> full interactive
+                         re-auth required; refresh loop will NOT help.
+  REFRESH_DEAD        -> refresh grant itself returns invalid_grant (pitfall 9)
+                         -> full interactive re-auth, or switch to a static API key.
+
+Usage:
+  python3 diagnose-oauth-mcp.py <server_name> [--mcp-url URL] [--token-endpoint URL] [--write]
+
+  <server_name> matches the files in $HERMES_HOME/mcp-tokens/<server>.json etc.
+  If --mcp-url / --token-endpoint are omitted, they're read from the token's
+  `resource` field and the AS .well-known metadata respectively.
+
+NEVER prints secret values — only lengths, scope, expiry, and HTTP status.
+"""
+import json, os, sys, time, argparse, urllib.request, urllib.error, urllib.parse
+
+UA = "python-httpx/0.27"  # CF blocks default urllib UA on many providers
+
+
+def _hermes_home():
+    # Prefer Hermes' own resolver (profile-safe); fall back to env then ~/.hermes.
+    try:
+        from hermes_constants import get_hermes_home
+        return str(get_hermes_home())
+    except Exception:
+        return os.environ.get("HERMES_HOME") or os.path.expanduser("~/.hermes")
+
+
+def _tokens_dir():
+    return os.path.join(_hermes_home(), "mcp-tokens")
+
+
+def _post(url, data=None, headers=None, form=False, timeout=30):
+    if form:
+        body = urllib.parse.urlencode(data).encode()
+    else:
+        body = json.dumps(data).encode() if data is not None else None
+    req = urllib.request.Request(url, data=body, method="POST")
+    for k, v in (headers or {}).items():
+        req.add_header(k, v)
+    req.add_header("User-Agent", UA)
+    try:
+        r = urllib.request.urlopen(req, timeout=timeout)
+        return r.status, dict(r.headers), r.read()
+    except urllib.error.HTTPError as e:
+        return e.code, dict(e.headers), e.read()
+
+
+def _get_json(url, timeout=20):
+    req = urllib.request.Request(url, headers={"User-Agent": UA})
+    return json.loads(urllib.request.urlopen(req, timeout=timeout).read())
+
+
+def _mcp_initialize(mcp_url, access_token):
+    status, hdrs, body = _post(
+        mcp_url,
+        data={
+            "jsonrpc": "2.0", "id": 1, "method": "initialize",
+            "params": {"protocolVersion": "2025-06-18", "capabilities": {},
+                       "clientInfo": {"name": "hermes-diag", "version": "1.0"}},
+        },
+        headers={
+            "Authorization": "Bearer " + access_token,
+            "Accept": "application/json, text/event-stream",
+            "Content-Type": "application/json",
+            "MCP-Protocol-Version": "2025-06-18",
+        },
+    )
+    txt = body[:400].decode(errors="replace")
+    ok = status == 200 and ("serverInfo" in txt or '"result"' in txt)
+    expired = "-32002" in txt or "Session expired" in txt or "invalid_token" in (hdrs.get("WWW-Authenticate") or "")
+    return ok, expired, status, txt
+
+
+def main():
+    ap = argparse.ArgumentParser()
+    ap.add_argument("server")
+    ap.add_argument("--mcp-url")
+    ap.add_argument("--token-endpoint")
+    ap.add_argument("--write", action="store_true", help="persist a working refreshed token")
+    args = ap.parse_args()
+
+    tdir = _tokens_dir()
+    tpath = os.path.join(tdir, args.server + ".json")
+    cpath = os.path.join(tdir, args.server + ".client.json")
+    tok = json.load(open(tpath))
+    client = json.load(open(cpath))
+
+    mcp_url = args.mcp_url or tok.get("resource")
+    if not mcp_url:
+        print("FATAL: no --mcp-url and no `resource` in token file"); sys.exit(2)
+    resource = tok.get("resource") or mcp_url
+
+    print(f"server={args.server} mcp_url={mcp_url}")
+    exp = tok.get("expires_at")
+    if isinstance(exp, (int, float)):
+        print(f"stored expires_at in {round((exp - time.time())/60)} min")
+
+    # Step 1: stored token
+    ok, expired, status, txt = _mcp_initialize(mcp_url, tok["access_token"])
+    print(f"[1] stored-token initialize -> HTTP {status} ok={ok} expired={expired}")
+    if ok:
+        print("BRANCH=TOKEN_OK  -> stored token works; 'not connected' is the breaker (7). Restart the gateway.")
+        return
+
+    # Step 2: refresh
+    if not tok.get("refresh_token"):
+        print("BRANCH=REFRESH_DEAD  -> no refresh_token present; full re-auth required.")
+        return
+    token_ep = args.token_endpoint
+    if not token_ep:
+        # derive AS metadata from the host root
+        host = urllib.parse.urlsplit(mcp_url)
+        as_url = f"{host.scheme}://{host.netloc}/.well-known/oauth-authorization-server"
+        try:
+            token_ep = _get_json(as_url)["token_endpoint"]
+        except Exception as e:
+            print(f"FATAL: cannot discover token_endpoint ({e}); pass --token-endpoint"); sys.exit(2)
+    print(f"[2] token_endpoint={token_ep}")
+
+    rstatus, _rh, rbody = _post(
+        token_ep, form=True,
+        data={"grant_type": "refresh_token", "refresh_token": tok["refresh_token"],
+              "client_id": client["client_id"], "resource": resource},
+        headers={"Content-Type": "application/x-www-form-urlencoded"},
+    )
+    if rstatus != 200:
+        print(f"[2] refresh -> HTTP {rstatus} {rbody[:200].decode(errors='replace')}")
+        print("BRANCH=REFRESH_DEAD  -> refresh grant rejected (9). Full re-auth or switch to static API key.")
+        return
+    j = json.loads(rbody)
+    new_at = j["access_token"]
+    print(f"[2] refresh -> HTTP 200 new_token_len={len(new_at)} scope={j.get('scope')} "
+          f"expires_in={j.get('expires_in')} rotated_refresh={bool(j.get('refresh_token'))}")
+
+    # Step 2b: smoke-test the freshly-minted token
+    ok2, expired2, status2, txt2 = _mcp_initialize(mcp_url, new_at)
+    print(f"[2b] new-token initialize -> HTTP {status2} ok={ok2} expired={expired2}")
+    if ok2:
+        if args.write:
+            new = dict(tok)
+            new.update({"access_token": new_at, "token_type": j.get("token_type", "Bearer"),
+                        "expires_in": j.get("expires_in", tok.get("expires_in")),
+                        "scope": j.get("scope", tok.get("scope")),
+                        "expires_at": time.time() + float(j.get("expires_in", 3600))})
+            if j.get("refresh_token"):
+                new["refresh_token"] = j["refresh_token"]
+            tmp = tpath + ".tmp"
+            open(tmp, "w").write(json.dumps(new, indent=2))
+            os.chmod(tmp, 0o600)
+            os.replace(tmp, tpath)
+            print(f"     wrote {tpath} (0600). NOW RESTART the gateway to clear the breaker.")
+        print("BRANCH=REFRESH_FIXED  -> refreshed token works. Persist (--write) + restart gateway.")
+        return
+
+    if expired2:
+        print("BRANCH=SESSION_REVOKED  -> refresh succeeds but new token STILL -32002 'Session expired' (10).")
+        print("     Refresh loop will NOT help. Full interactive authorization_code re-auth required.")
+        print("     For an unattended gateway, prefer a static Personal API key instead.")
+        return
+    print(f"BRANCH=UNKNOWN  -> new token failed for a non-session reason: {txt2[:200]}")
+
+
+if __name__ == "__main__":
+    main()