Anthropic

Claude directly, over the Messages API — not a chat-completions shim. The system turn is a top-level field and the reply arrives as content blocks, so this provider carries its own HTTP shape while sharing every prompt with the rest.

Good for

• Teams that want Claude and already have an Anthropic account.
• Instruction-following on the classifier's strict JSON contract, which is the part that decides whether the AI layer contributes anything at all — a reply that is not the demanded shape is discarded, not salvaged.
• A cheap default: the built-in model is a Haiku-class one, chosen because classification is high-volume and low-difficulty.

Configuration

[llm]
provider = "anthropic"

[llm.providers.anthropic]
api_key_env = "ANTHROPIC_API_KEY"        # required
model = "claude-haiku-4-5-20251001"      # set this explicitly
max_tokens = 512
# api_base = "https://api.anthropic.com"

export ANTHROPIC_API_KEY=sk-ant-...

Key	Required	Default	Notes
api_key_env	✅	—	Env var holding the key
model	—	claude-haiku-4-5-20251001	Any Claude model the account can reach
api_base	—	https://api.anthropic.com	Trailing slash trimmed
max_tokens	—	512	Cap on the reply

The pinned API version (anthropic-version: 2023-06-01) is sent on every request and is not configurable — pinning it is what keeps a server-side API change from silently altering verdicts.

note

api_base has no /v1 here

Unlike the OpenAI-shaped providers, the Anthropic default base is https://api.anthropic.com — the version segment is part of the request path the provider builds, not of api_base.

Choosing a model

Want	Try
Cheapest sensible default	a Haiku-class model (the built-in default)
Better judgement on borderline messages	a Sonnet-class model
Deeper reasoning than this job needs	an Opus-class model — almost certainly overkill

Anthropic model ids are dated (claude-haiku-4-5-20251001). That is a feature: pinning the date means today's verdicts are still today's verdicts next month. Set model explicitly and update it on your own schedule.

Both jobs are small. Classification is a one-line JSON verdict; a rewrite is one short message, not an essay. There is no reason to reach for a large model unless you have measured the borderline cases and found the small one wanting.

list_models returns an empty list for this provider — the dashboard's model picker has no catalog to browse here, so type the id. If you want a live picker, put OpenRouter first in the chain.

The strict-JSON contract

The classifier is given the exact reply shape verbatim in the system prompt, so the model has no format latitude:

{"verdict":"clear"|"low_context","confidence":0.0-1.0,
 "missing":["details"|"link"|"recipients"],
 "reasons":["short human sentence for the author"]}

Anything else is an error, and an error is read as "no opinion" — the heuristic verdict stands. The parser tolerates fenced or prose-wrapped JSON, but it will not invent a verdict out of a reply that does not contain one.

Cost and privacy

• Only ambiguous messages are sent. Clear passes and textbook fails never leave the process. Under sensitivity = "conservative" nothing is sent.
• What is sent: the message text and the team charter; for a rewrite, resolved context items too.
• Replies are capped at max_tokens (512 by default).
• Requests time out after 30 seconds and fall through to the next provider.
• If message content must not reach a vendor at all, use Ollama.

Errors and secrets

Non-2xx bodies are quoted as a capped excerpt (300 characters) so an HTML error page cannot flood the logs, and the API key never appears in an error — it exists only in the x-api-key header.

A 401 is the key; a 404 on a valid key is usually a model id the account cannot reach, often a stale date suffix; a 429 is what the fallback chain exists for.