Connect Claude to Caryvane

Caryvane's MCP server is one HTTPS endpoint; Claude needs its URL and your key. Five minutes, then "which backups failed last night?" is a question rather than a report.

1. Create a key #

Console → your name → AI access keys → New Key. Name it "Claude", choose the tier — Read only for questions, Read + control to let it start and cancel runs, Read + control + edit to let it create jobs and mount drives — and copy the key; it is shown once. Details of the tiers on AI assistants & API.

2a. Claude Code #

From the terminal, in the project you want it available in (or with --scope user for everywhere):

claude mcp add --transport http caryvane https://console.caryvane.com/mcp/rpc \
  --header "Authorization: Bearer <key>"

Or write .mcp.json in the project:

{
  "mcpServers": {
    "caryvane": {
      "type": "http",
      "url": "https://console.caryvane.com/mcp/rpc",
      "headers": { "Authorization": "Bearer <key>" }
    }
  }
}

Keep the file out of source control — it holds the key. Restart Claude Code; /mcp lists the server and its tools.

2b. Claude Desktop and claude.ai #

This one trips people up, because the obvious-looking button is the wrong one. Caryvane authenticates with a key you paste as a header, not with a sign-in, so there is nothing to sign in to and no window will pop up. Work down the panel the console showed you when it created the key — it has a copy button beside each of the three values.

Settings → Connectors → Add custom connector.
Name it Caryvane.
Paste the server URL: https://console.caryvane.com/mcp/rpc (use your own console address if it differs).
Do not sign in, and leave the OAuth boxes empty. There is no client ID and no client secret. If you click Connect expecting a login, it will fail.
Add a custom header. Name Authorization. Value Bearer followed by your key, one space between the two, nothing else. The console offers this already assembled — copying Header value saves typing it.
Save and enable it for the conversation. The Caryvane tools appear in the tools menu.

The single most common mistake is pasting the key on its own into the header value. The word Bearer and the space are both part of it.

2c. ChatGPT #

ChatGPT reaches the same endpoint the same way, under Settings → Connectors → Create (on some plans this sits behind developer mode). Give it the name, the same /mcp/rpc URL, and for authentication choose the access-token or API-key option rather than OAuth, then paste your key. ChatGPT sends it as Authorization: Bearer <key>, which is exactly what Caryvane expects. Menu wording moves between releases; the three values never do.

Caryvane does not offer OAuth, so a client that insists on it cannot connect. If yours does, use the plain REST surface on AI assistants & API instead — same tools, same tiers, ordinary HTTP calls with the same bearer header.

3. Use it #

The tool descriptions tell Claude what each does and what to check first, so plain requests work:

"Which jobs failed or warned in the last 24 hours, and why?" — get_dashboard, list_jobs, get_job_runs.
"Is HHSRV1 online and what version is it on?" — list_agents.
"Show me the agent log for BTBS2 around 2 am." — get_agent_logs.
"Run the Henderson nightly job now." — run_job (Control tier).
"Create a nightly backup of D:\Shares on HHSRV1 to the Henderson bucket, keep 5 versions." — create_job (Edit tier).
"Make me an enrolment token for Modobloc." — create_enrollment_token (Full tier).

Claude sees exactly what you see. It cannot download file contents, delete files from storage or move them; browsing returns names and dates only.

If it does not connect #

401 — the key is wrong, revoked, or pasted with a trailing space.
Tools missing after a change on our side — reconnect the server (Claude caches the tool list at connect time).
"Not allowed" on a tool — the key's tier is too low; issue a higher-tier key.