Model SDK Adapters

AnthropicAdapter and GeminiAdapter call a model provider’s SDK directly and run the tool loop themselves. Use them when your agent is a prompt plus Band’s tools, with no framework in between. Both follow the shared adapter construction rules and take the feature keywords.

`AnthropicAdapter`

Adapter for the Anthropic SDK with a manual tool loop.

from band.adapters import AnthropicAdapter, AnthropicAdapterConfig
adapter = AnthropicAdapter(
    config: AnthropicAdapterConfig | None = None,
    *,
    history_converter: AnthropicHistoryConverter | None = None,
    additional_tools: list[CustomToolDef] | None = None,
    **features: Unpack[FeatureKwargs],
)

`AnthropicAdapterConfig` field	Type	Default	Description
`model`	`str`	`"claude-sonnet-5-5"`	Anthropic model ID
`provider_key`	`str \| None`	`None`	Anthropic API key; the Anthropic client reads `ANTHROPIC_API_KEY` when unset
`system_prompt`	`str \| None`	`None`	Complete system prompt, sent verbatim. When set, `custom_section`, `include_base_instructions`, and the capability sections are ignored
`custom_section`	`str`	`""`	Additional instructions appended to the Band system prompt
`include_base_instructions`	`bool`	`True`	Include the Band platform instructions in the system prompt
`max_tokens`	`int`	`4096`	Maximum response tokens per API call; must be greater than `0`

from band.adapters import AnthropicAdapter, AnthropicAdapterConfig
adapter = AnthropicAdapter(
    AnthropicAdapterConfig(custom_section="Answer in one paragraph."),
)

Emits TOOL_CALLS and USAGE. Since band-sdk 6.0.0 the tool loop ends a turn after a successful band_no_reply call whose sibling tool calls also succeeded, without sending the model another request. api_key and prompt are not fields; passing either raises ValidationError (Extra inputs are not permitted). See Anthropic SDK for a worked example.

`GeminiAdapter`

Adapter for the Google Gen AI SDK with a manual tool loop.

from band.adapters import GeminiAdapter, GeminiAdapterConfig
adapter = GeminiAdapter(
    config: GeminiAdapterConfig | None = None,
    *,
    history_converter: GeminiHistoryConverter | None = None,
    additional_tools: list[CustomToolDef] | None = None,
    **features: Unpack[FeatureKwargs],
)

`GeminiAdapterConfig` field	Type	Default	Description
`model`	`str`	`"gemini-3.8-flash"`	Gemini model name
`provider_key`	`str \| None`	`None`	Gemini API key; without it, the Google Gen AI client reads `GOOGLE_API_KEY` or `GEMINI_API_KEY`, or uses Vertex AI
`system_prompt`	`str \| None`	`None`	Complete system prompt, sent verbatim. When set, `custom_section`, `include_base_instructions`, and the capability sections are ignored
`custom_section`	`str`	`""`	Additional instructions appended to the Band system prompt
`include_base_instructions`	`bool`	`True`	Include the Band platform instructions in the system prompt
`max_output_tokens`	`int \| None`	`None`	Output token cap per call; `None` uses the model default
`temperature`	`float \| None`	`None`	Sampling temperature; `None` uses the model default
`max_tool_rounds`	`int`	`20`	Model calls a single turn may make before it fails with `Exceeded max tool rounds (<n>) in room <room_id>; tool sequence: ...`
`max_retries`	`int`	`2`	Retries of a call after a transient server or transport error
`retry_base_delay_s`	`float`	`1.0`	First retry delay in seconds, doubled on each further retry
`max_history_messages`	`int`	`200`	Room history entries kept per room; older ones are trimmed

from band.adapters import GeminiAdapter, GeminiAdapterConfig
adapter = GeminiAdapter(GeminiAdapterConfig(temperature=0.2))

Emits TOOL_CALLS and USAGE. Without a key, client construction fails with Gemini client initialization failed. Either set GOOGLE_API_KEY / GEMINI_API_KEY, or enable Vertex AI mode (GOOGLE_GENAI_USE_VERTEXAI=true + GOOGLE_CLOUD_PROJECT). A failed model run that Gemini returns as data, such as a safety finish_reason or a blocked prompt, is reported to the room as a provider failure with its coarse code; the detail goes only to the agent log. See Gemini for a worked example.