os-mcp

OS MCP — Railway Deploy

This skill deploys the Relay MCP v2 server to the user's own Railway account. The MCP source is bundled at ${CLAUDE_PLUGIN_ROOT}/skills/os-mcp/reference/relay-mcp-server/ — copy from there, do not fetch from anywhere else. The deployed server lets Claude clients (Claude Code or claude.ai) read/write the user's Obsidian vault by talking to the Relay.md sync protocol.

What the user provides

One thing: a Railway account token — created at https://railway.com/account/tokens.

That's it. After deploy, the user signs in with their Relay.md credentials via OAuth; the MCP then auto-discovers their relay vault and the folders inside it via PocketBase. No vault GUIDs, no folder maps.

Values set automatically by the skill

| Var | Source | Value | |-----|--------|-------| | RELAY_API_URL | preset | https://api.system3.md | | PB_AUTH_URL | preset | https://auth.system3.md | | PB_COLLECTION | preset | users | | DATA_DIR | preset | /data | | PORT | preset | 3000 | | JWT_SECRET | auto-generated | openssl rand -hex 32 | | STATIC_MCP_BEARER | auto-generated | openssl rand -hex 24 | | ALLOWED_EMAILS | preset | blank (any authenticated Relay.md user) | | RELAY_AUTH_TOKEN | preset | blank (OAuth-only; no static fallback) | | RELAY_ID | not set | auto-discovered at runtime per user | | PUBLIC_URL | derived | from railway domain after first deploy |

Save the generated STATIC_MCP_BEARER in the chat for the user — they'd use it for any CLI/script access bypassing OAuth.

If the user has multiple Relay vaults, the MCP auto-picks the first one. They can call the vault_relays tool after connecting to see which is active and pin a different one by setting RELAY_ID in railway variables.

Workflow

Execute in order. Confirm each step before moving on.

Step 1 — Verify Railway CLI

railway --version

If missing:

• macOS: brew install railway
• Anywhere: npm i -g @railway/cli

Re-run to confirm.

Step 2 — Get Railway token from the user

Tell the user:

Open https://railway.com/account/tokens, click Create Token, name it os-mcp-deploy, copy the value, and paste it here.

When pasted:

export RAILWAY_API_TOKEN="<paste>"
railway whoami

railway whoami should print the user's Railway email. If it errors, the token is wrong — ask them to regenerate.

Step 3 — Deploy the bundled source

cd "${CLAUDE_PLUGIN_ROOT}/skills/os-mcp/reference/relay-mcp-server"

# Create a new Railway project (interactive — let user pick name)
railway init

# Persistent volume for clients.json + sessions.enc
railway volume add --mount-path /data

# Upload source, build via Dockerfile, deploy
railway up

Wait for the deploy. Tail logs if useful: railway logs.

Step 4 — Wire env vars

Generate the secrets:

JWT_SECRET=$(openssl rand -hex 32)
STATIC_MCP_BEARER=$(openssl rand -hex 24)

Set everything in one shot (note: RELAY_ID is intentionally NOT set — it auto-discovers):

railway variables \
  --set "RELAY_API_URL=https://api.system3.md" \
  --set "PB_AUTH_URL=https://auth.system3.md" \
  --set "PB_COLLECTION=users" \
  --set "DATA_DIR=/data" \
  --set "PORT=3000" \
  --set "JWT_SECRET=$JWT_SECRET" \
  --set "STATIC_MCP_BEARER=$STATIC_MCP_BEARER"

Service auto-redeploys after --set. Wait for it to come back up.

Step 5 — Set PUBLIC_URL and final redeploy

railway domain                      # assigns + prints public domain

Copy the URL, then:

railway variables --set "PUBLIC_URL=https://<the-domain>"
railway up                          # final redeploy so OAuth issuer/audience match

Step 6 — Verify

curl https://<the-domain>/health

Expect 200 OK with JSON {"status":"ok","version":"2.0.0",...}.

Step 7 — Connect from Claude

Print these for the user:

Claude Code:

claude mcp add --transport http vault https://<the-domain>/mcp

Claude.ai (web): Settings → Connectors → Add custom connector → paste https://<the-domain>/mcp. Sign in with their Relay.md credentials when the OAuth window opens.

Also tell them: their STATIC_MCP_BEARER is <value> — keep it for any CLI/script access that needs to bypass OAuth.

Step 8 — Test discovery

In a fresh Claude conversation, ask Claude to call vault_relays (or have the user say "list my Relay vaults"). Expect at least one vault, with * next to the active one.

Then call vault_folders — expect a non-empty list of folder names. If empty, the user may not have folder access in that relay.

If the user has multiple relays and the wrong one is active, pin the right one:

railway variables --set "RELAY_ID=<the-vault-guid-from-vault_relays>"

(One redeploy and the next OAuth session will use that relay.)

Troubleshooting

railway whoami fails after pasting Railway token. Token expired or wrong scope. Regenerate at https://railway.com/account/tokens.

Build fails with TypeScript errors. Don't modify the bundled package.json / tsconfig.json. Re-copy fresh from ${CLAUDE_PLUGIN_ROOT}/skills/os-mcp/reference/relay-mcp-server/.

/health returns 5xx. railway logs will name the missing env var (assertRequiredConfig lists what's missing on boot).

vault_relays returns "No relays accessible to this user". The signed-in Relay.md account doesn't own or belong to any relay. Have them create one in the Relay.md Obsidian plugin first.

vault_folders returns empty but vault_relays shows the vault. Folders are stored in the shared_folders PocketBase collection — the user may not have any folders in that relay yet, or the active relay isn't the one they meant. Check vault_relays output for the active marker.

OAuth sign-in loops. PUBLIC_URL doesn't match the actual Railway domain. Re-run Step 5.

Want to gate access to specific emails. railway variables --set "ALLOWED_EMAILS=email1@x.com,email2@x.com" then redeploy.

Want to start over. railway down to tear down, then re-run from Step 3.