Consumer adapters

Karta’s native session API is one way to drive an agent. The consumer adapters are another: they expose the same turn under three familiar wire formats, so a client built for OpenAI or Anthropic can talk to a Karta agent after you point it at the agent URL and use the right credential.

OpenAI Responses

POST /{org}/{agent}/v1/responses - the OpenAI Responses API shape, with previous_response_id continuation.

OpenAI Chat Completions

POST /{org}/{agent}/v1/chat/completions - the classic messages-array shape.

Anthropic Managed Agents

Event-sourced managed-agent sessions with a streamed agent.* event log.

Pick an adapter

Adapter	Best for	Continuity	Streaming shape	Avoid when
OpenAI Responses	New OpenAI-style clients that can use Responses.	Continue with `previous_response_id` when `store` is set.	Named SSE events such as `response.output_text.delta`.	Your client only supports Chat Completions.
OpenAI Chat Completions	Existing clients that require the classic messages-array API.	Stateless: each call runs the latest user message on a fresh session.	Bare `data:` chunks ending with `[DONE]`.	You need durable continuation, approval details, files, or event-log replay.
Anthropic Managed Agents	UIs that post events, replay logs, upload files, confirm tools, or resume sessions.	Session id plus ordered `agent.*` event log.	`agent.`, `session.`, and `warm` SSE events.	You only need a single request/response call.
Native sessions	Backends that want Karta’s direct session/message/input API.	Session id and Karta typed events.	Karta event names from Streaming events.	You need an OpenAI- or Anthropic-shaped SDK without adapter code.

Why adapters

The point of a fixed session protocol is that one client talks to any agent. The adapters extend that idea to clients you did not write for Karta: point an existing OpenAI or Anthropic SDK at your agent URL and it works, because Karta maps its typed event stream onto each format’s wire shape.

        +- /responses          -> OpenAI Responses shape
Karta   +- /chat/completions   -> OpenAI Chat Completions shape
turn -->+- /managed-agents     -> Anthropic Managed Agents events
        +- /sessions/.../messages -> native Karta shape
        (one turn, four projections)

Auth & scope

All adapters live on the agent route family (/{org}/{agent}/v1/... on agent.karta.sh), so they accept either a session token (the browser-safe credential) or a kt_live_... API key. They enforce the same budget gate (402) and agent pinning as the native routes.

Streaming

Every adapter supports stream: true and emits SSE in that format’s event vocabulary - response.output_text.delta for Responses, chunk for Chat Completions, agent.* for Managed Agents - rather than Karta’s native event names. Pick the adapter that matches your client and its streaming expectations.

These adapters are a compatibility surface, not a re-implementation of the providers’ full APIs. They cover the conversational turn - input, streamed output, usage. Provider-specific features beyond that are not guaranteed.

Overview

Sessions