@irisrun/provider-anthropic

A vendor-neutral, replay-safe model adapter. The Anthropic model_call performer sits behind Iris's portable model port: the agent names a model, not a vendor, and because each reply is captured as a journaled effect, a recorded session replays byte-identically from its journal — no provider is baked into the core, and swapping one out never changes a past session's replay. (That's faithful record-replay of the captured reply, not a claim that the model itself is deterministic.)

What it is

A direct adapter for the Anthropic Messages API over the built-in fetch — zero runtime dependencies. It exposes a buffered performer (anthropicModelPerformer) and a streaming one (anthropicStreamingModelPerformer) that emit token deltas while holding the content === join(deltas) reconcile invariant. The real adapter calls Anthropic only when ANTHROPIC_API_KEY is set; the test suite runs it against a fake fetch / fake model, so nothing in the suite touches the network.

Use it

Selected automatically from an anthropic/… model prefix in your Agentfile — set the prefix once and iris run / chat / serve / deploy all follow it:

# agent.yaml
model: anthropic/claude-x

This adapter speaks the Anthropic Messages protocol, not just Anthropic itself: point it at any compatible endpoint with --base-url (or IRIS_MODEL_BASE_URL) on iris run / serve / chat — Bedrock and Vertex Claude among them. Which endpoints are replay-safe vs known-divergent is a conformance-tested matrix in @irisrun/provider-compat (iris providers --matrix).

See docs/Models & providers for the model port, the shared conformance suite, the compatibility matrix, and how to add a third provider.

Part of Iris — own, portable, verifiable state.