AI assistants & API

Caryvane exposes an MCP server and a matching REST API, authenticated by personal access keys. An assistant with a key can answer "which backups failed last night?", run a job, enrol a machine, or set up a customer's storage — bounded by the key's tier and by what the user who issued it is allowed to see.

Create a key #

1. Console → your name (top right) → AI access keys → New Key. Give it a name you will recognise when revoking it.
2. Choose the tier. Least privilege that does the job — a key for a reporting assistant should be Read only.
3. Copy the key. It is shown once, at creation, and never again. Revoke and reissue if it is lost.

A key sees exactly what its owner sees: an MSP admin's key sees that MSP's customers, a customer user's key sees one customer. Every action a key takes is written to the audit log under the owner's name. Keys are revoked from the same dialog and stop working at once.

Connect Claude #

The MCP server is a single HTTP endpoint, https://console.caryvane.com/mcp/rpc, with the key as a bearer token. In Claude Code, add to .mcp.json in the project (or ~/.claude.json for every project):

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

Or from the terminal: claude mcp add --transport http caryvane https://console.caryvane.com/mcp/rpc --header "Authorization: Bearer <key>". Claude Desktop and claude.ai take the same URL and header under Settings → Connectors → Add custom connector. After connecting, ask it to list the tools; the description of each is written for the assistant, so it knows to check list_agents before create_job.

Connect VS Code (Copilot agent mode) #

Add to .vscode/mcp.json (VS Code prompts for the key once and stores it in its secret storage):

{
  "inputs": [
    { "id": "caryvane-key", "type": "promptString", "password": true, "description": "Caryvane access key" }
  ],
  "servers": {
    "caryvane": {
      "type": "http",
      "url": "https://console.caryvane.com/mcp/rpc",
      "headers": { "Authorization": "Bearer ${input:caryvane-key}" }
    }
  }
}

The tools #

Names are the same in MCP and in the REST API below. Tier is the minimum key tier.

The REST API #

The same operations as plain JSON over HTTPS, for scripts, RMM integrations and any assistant that speaks HTTP rather than MCP. Same key, same header: Authorization: Bearer <key>. Responses are JSON; errors are standard status codes with a message.

curl -s https://console.caryvane.com/mcp/jobs?take=50 \
  -H "Authorization: Bearer <key>"

curl -s -X POST https://console.caryvane.com/mcp/jobs/14/run \
  -H "Authorization: Bearer <key>"

Parameters are the tool's arguments as JSON in the body (POST) or the query string (GET). The MCP tool descriptions, which the assistant sees, are the authoritative field reference; an OpenAPI document for building a ChatGPT Custom GPT Action is on the list but not published yet.

What an assistant cannot do #

By design there is no tool that downloads a file's contents, deletes files from storage, or moves them. Browsing returns metadata only. Restores are done by people in the console. The one operation that removes data — an Archiving job with source removal switched on — is created like any other job, with the gate off unless explicitly set, and every file it removes is recorded in the catalogue first.