Skip to content

Anthropic (Claude)

  1. Go to console.anthropic.com.
  2. Settings → API Keys → Create key.
  3. Copy the sk-ant-... token.

Anthropic has no free tier. You’ll need to add a payment method and fund your workspace with at least $5 to make calls.

Terminal window
ANTHROPIC_API_KEY=sk-ant-...

Restart the backend container. The key is validated on start; a bad key shows up as [env:anthropic] Invalid API key in the logs.

Settings → AI providers → Anthropic → Paste key → Save. The UI runs a validation call (/v1/models) before accepting. Saved keys are encrypted before storage.

RoleDefault modelWhy
Planner (chief-of-staff)claude-sonnet-4-6Best plan-quality per dollar.
Critique + reviseSame as plannerSame context, same model keeps consistency.
Specialists (BA, architect, QA)claude-haiku-4-5Fast + cheap for routine work.
Long transcripts (summarisation)claude-sonnet-4-6Handles 200k-token windows.

Override any role via MODEL_<ROLE>_ANTHROPIC env vars. See Model registry.

Anthropic’s rate limits are per-workspace and tier-gated:

  • Tier 1 (new account, <$100 spent): 50 req/min, ~1M tokens/min.
  • Tier 2 ($100+): 1k req/min, 2M+ tokens/min.
  • Tier 3+: enterprise, much higher.

Workforce0 respects 429 responses by backing off and retrying. Persistent 429s degrade this provider in the AI Council for 60s.

Based on Anthropic’s published pricing at time of writing:

ModelInput $/1MOutput $/1MTypical brief cost
Claude Sonnet 4.6$3.00$15.00~$0.10
Claude Haiku 4.5$0.80$4.00~$0.02

Prompt caching is enabled by default — see Model registry.

If you have a Claude Code (or Pro) subscription on your local machine, the agent daemon can use it for code-gen via the CLI, not the API. This is handled separately from the BYOK Anthropic key — see Agent daemon.

The subscription key is never transmitted over the network. It stays on your machine.

The provider page shows red. Calls fall back to other Council members. Fix: rotate the key.

Anthropic doesn’t throttle on spend (paid tier). Workforce0’s PLANNER_MONTHLY_BUDGET_TOKENS gates do — see Cost caps.

Some regions have occasional 503s. The retry loop handles it. If sustained, switch to OpenAI or Gemini until Anthropic recovers.

From Anthropic’s current policy:

  • API inputs and outputs are NOT used to train models by default.
  • Retention: minimum necessary for providing the service.
  • Workforce0 doesn’t opt in to anything extra.

Read the current policy at anthropic.com/legal.