@tailored-ai/provider-anthropic

Anthropic Messages API provider for Tailored AI agents — Claude models with prompt caching, beta headers, streaming, and model discovery.

This plugin provides the anthropic provider id (#236 moved it out of core). Existing providers.anthropic config keeps working, and gains the features below. On older cores that still ship a minimal built-in, the plugin supersedes it — the one-line "Replacing existing entry" notice at startup is expected.

Install

tai plugin install @tailored-ai/provider-anthropic

Configure

plugins:
  - "@tailored-ai/provider-anthropic"

providers:
  anthropic:
    apiKey: "${ANTHROPIC_API_KEY}"     # https://console.anthropic.com
    defaultModel: "claude-haiku-4-5"
    promptCaching: true                # optional, recommended for agents

agent:
  defaultProvider: anthropic

| Field | Required | Notes | |---|---|---| | apiKey | yes | Use ${ANTHROPIC_API_KEY} to read from the environment. | | defaultModel | yes | A Claude model id, e.g. claude-haiku-4-5. tai edit lists them via listModels. | | promptCaching | no | Adds ephemeral cache breakpoints to the system prompt and tool definitions. Agent loops re-send both every iteration, so cache hits cut input cost (cached reads are ~10% of base price) and latency. Default false. | | defaultMaxTokens | no | max_tokens when the agent doesn't set one (the API requires it). Default 4096. | | version | no | anthropic-version header. Default 2023-06-01. | | betas | no | List of beta flags, sent as the anthropic-beta header (e.g. context-1m-2025-08-07). | | baseUrl | no | Default https://api.anthropic.com. Override for proxies. |

Extra request fields

Anthropic-specific parameters the TAI contract doesn't model (extended thinking, top_k, …) flow through ChatParams.extra straight into the request body. Surfaces that expose a providerExtra config bag (briefing, suggestions) can use it today; the main agent loop doesn't set extra per agent yet.

Usage accounting

Cache writes and reads are summed into usage.input, so token counts reflect what the API actually processed rather than silently excluding cached tokens.

Development

pnpm --filter @tailored-ai/provider-anthropic run build
pnpm --filter @tailored-ai/provider-anthropic run test

License

MIT