Client Tool Compatibility

The OpenAI Responses API defines several client tool types (tool_search, local_shell, custom, and more). This page lists their compatibility across upstream protocols, and which models have a Responses-native upstream.

Responses-native models

These models have at least one supplier route that natively speaks the Responses API. When a request is routed to such a route, server-executed tools (web_search, mcp, the server variant of tool_search, etc.) are carried by the upstream itself; when routed elsewhere, the matrix below applies.

responses_native only guarantees that at least one such route exists for the model — it does not mean this particular call will be routed there or is exempt from the matrix below. Routing is decided by the platform.

To see which models are currently Responses-native, query the model catalog endpoint (GET /api/v1/spus, field responses_native).

Tool type compatibility matrix

Tool typeWhen routed to a Responses-native routeOtherwise
functionSupportedSupported
custom (e.g. Codex exec)SupportedAdapted (Gemini routes pending)
local_shellSupportedAdapted
namespace (tool groups)Supported (flattened)Same
tool_search · execution:"client"SupportedAdapted for OpenAI/Anthropic protocols (Gemini unverified)
tool_search · server-side searchSupportedNot supported: 400 with reason=server_executed
shell (local) / apply_patchSupportedReturns 400 with reason=not_adapted
computer / computer_use_previewSupportedPlanned (needs multimodal tool output)
mcp / web_search / file_search / code_interpreter / image_generationSupportedNot supported: 400 with reason=server_executed

Status reflects the current production release and is updated on deploy; rejected requests state the reason and available alternatives in the error message.

Anthropic output and thinking controls

For Messages requests routed to an Anthropic upstream, see the Anthropic field reference for the supported output_config and thinking shapes.

See current Responses-native models · Find error codes