Setting up Claude Code

Claude Code is 🏰 Camelot AI's default Agent CLI, and the only one that's fully integrated and tested today — start here unless you already know you need Codex.

1. Get an API key

Either of these works:

Anthropic Console API key — an sk-ant-... key from the Anthropic Console.
Claude Code OAuth token — if you're on a Claude subscription plan that supports the CLI, run claude setup-token locally and copy the resulting sk-ant-oat... token. This is billed against your subscription instead of pay-as-you-go API usage.

Camelot detects which one you pasted by its prefix, so either format just works — see how the key reaches the runner below.

2. Add it as a Credential

Go to /profile (any signed-in user, not admin-only) and add a Credential:

Kind: claude_api_key
Name: anything memorable, e.g. anthropic
Value: the key or token from step 1

Credentials are encrypted at rest and shipped securely to the runner container that executes your tasks.

Why either key format "just works"

The claude_api_key credential is mapped to an environment variable before the runner starts:

A plain sk-ant-... API key becomes ANTHROPIC_API_KEY.
An sk-ant-oat... OAuth token becomes CLAUDE_CODE_OAUTH_TOKEN instead, with ANTHROPIC_API_KEY explicitly cleared so a stale key baked into the runner image can't take precedence.

You don't need to pick which env var to use — pasting either value into the same claude_api_key credential routes it correctly.

3. Pick Claude Code on a task

When creating a task on the board, the CLI Agent dropdown lets you choose which Agent CLI template runs it. Pick Claude Code — it's pre-seeded, so there's nothing to configure per-project beyond having the credential from step 2.

That credential is a hard prerequisite, not a suggestion: until the claude_api_key from step 2 exists, Claude Code — API key absent is greyed out in the dropdown and the form refuses the pick, rather than dispatching a run that would fail with a bare 401 minutes later inside the runner.

The board also needs somewhere to run the task: with no project at all, / sends you to /projects instead of opening the form.

4. Model discovery

The Model dropdown is resolved live, per user: Camelot calls GET https://api.anthropic.com/v1/models with the credential that will actually run the task and offers the claude-* ids that come back. No tokens are billed, and either key format works — an sk-ant-api… key goes on x-api-key, an sk-ant-oat… OAuth token on Authorization: Bearer, the same split the runner uses.

Whose key is used depends on where you are:

the New Task modal uses yours — you become the task's creator;
an existing task page uses the creator's, even when somebody else is looking at it, because the runner mounts the creator's key.

Each answer is cached for an hour (a minute, if the call failed), keyed to the user and to the key itself, so rotating or deleting the credential takes effect immediately.

When discovery can't answer — no key stored, the probe disabled, Anthropic unreachable or refusing the key — the dropdown falls back to the template's pinned available_models. Nothing about the page changes; you just see the offline list.

5. Admin: reviewing the template

Workspace admins can review or tune the Claude Code template at /agents (admin-only). The seeded defaults — executable claude, models claude-opus-5 / claude-sonnet-5 / claude-haiku-4-5-20251001, default claude-sonnet-5 — rarely need touching; those models are now the offline fallback for the dropdown rather than the whole story. Model discovery probe holds the JSON that drives discovery (clear it to {} to turn discovery off), and the per-row Check models action shows what your own key is entitled to, with Pin these to copy that into the fallback list. Nothing rewrites available_models automatically.

6. Self-hosted: runner image

Docker / Swarm backends run Claude Code inside the project's runner container. The prebuilt ghcr.io/t0ha/camelot-runner-claude:latest image already has it installed via npm install -g @anthropic-ai/claude-code.
Local dev (LocalPort backend) runs the CLI directly on the host running Camelot instead of in a container, so claude must be installed and on PATH there too.

Troubleshooting

401 / auth errors — usually means the wrong credential kind was used (double-check it's claude_api_key, not openai_api_key) or the key/token has expired.
Task pauses instead of erroring — Camelot recognizes a provider usage-limit rejection and pauses the task (and any other queued task on the same credential) instead of retrying or erroring. It resumes automatically once the limit resets, or via "Resume now" on the task.