@tailored-ai/provider-anthropic

Claude models via the Messages API: prompt caching, beta headers, configurable API version, extra-field passthrough, streaming, and model discovery.

This plugin is the home of the anthropic provider id (#236 moved it out of core) — your existing providers.anthropic block keeps working. On older cores that still ship a minimal built-in, the plugin supersedes it; the one-line "Replacing existing entry" startup notice is expected there.

Install

bash

tai plugin install @tailored-ai/provider-anthropic

Configure

yaml

providers:
  anthropic:
    apiKey: "${ANTHROPIC_API_KEY}"
    defaultModel: "claude-haiku-4-5"
    promptCaching: true

agent:
  defaultProvider: anthropic

Field	Required	Notes
`apiKey`	yes	Use `${ANTHROPIC_API_KEY}` to read from the environment.
`defaultModel`	yes	A Claude model id. `tai edit` lists them via `listModels`.
`promptCaching`	no	Cache breakpoints on the system prompt + tool definitions. Recommended for agents — both are re-sent every loop iteration, and cached reads cost ~10% of base input price.
`defaultMaxTokens`	no	`max_tokens` fallback (the API requires one). Default `4096`.
`version`	no	`anthropic-version` header. Default `2023-06-01`.
`betas`	no	Beta flags for the `anthropic-beta` header.
`baseUrl`	no	Default `https://api.anthropic.com`.

Anthropic-specific parameters

Fields the TAI provider contract doesn't model (extended thinking, top_k, …) pass through ChatParams.extra into the request body. Surfaces with a providerExtra config bag (briefing, suggestions) can use this today; per-agent extra isn't wired into the main loop yet.

Usage accounting

Cache writes and reads are summed into usage.input — token counts reflect what the API processed, with no silent exclusion of cached tokens.