Claude Code can operate Orca end to end: install the CLI, log in without a browser on the machine, create agents, run them in the cloud, and follow their output. Two setup routes, pick one.
Route 1: the Orca plugin (recommended)
The plugin ships the use-orca skill (which teaches Claude the golden paths) and auto-registers the Orca MCP server for every session:
Route 2: MCP server only
If you already have the orca CLI installed and signed in:
First run
Inside a Claude Code session, ask for anything Orca-shaped, for example: “create an Orca agent that triages my support inbox and run it against these three sample tickets.”
Claude will walk the loop itself:
orca doctor --json to check the install, then curl -fsSL https://orcapods.ai/install.sh | sh if the CLI is missing.
orca login: in an agent context this automatically uses the device flow. Claude relays a one-time code and a URL; you open the URL on any device, check the code matches, and approve. The CLI picks up its key by polling; no key ever passes through the chat.
whoami, create_agent, run_agent, wait_for_run over MCP (or the --json CLI equivalents) to do the work.
A worked transcript
Notes
- The MCP server resolves credentials lazily (
ORCA_API_KEY, then the API key stored in the selected CLI context; there is no CLI flag that supplies the key itself, --context only selects which stored context to use), so registering it before the first login is fine; tools answer with the exact fix until you sign in.
- Everything not covered by a dedicated tool is reachable through the
api_request tool, documented by the live spec at the orca://openapi resource.
- Headless auth details, including CI: Headless authentication.
Next
Claude Code is wired up. Build it with your coding agent is the workflow from here: what to ask for, which of the two files each rule belongs in, and the loop that takes an agent from first draft to something you would put in front of a customer. Last modified on September 6, 2026