Skip to main content
POST /api/v1/chat/fusion sends the same conversation to two or three distinct models, then synthesizes their answers. It returns one non-streaming JSON result.

​
Authentication and billing

Use an organization API key as Authorization: Bearer or X-API-Key, or an OAuth bearer with cli:inference. API credentials do not require browser cookies or CSRF headers. Calls are bound to the credential’s organization. Fusion requires a paid plan with Fusion access. A prepaid balance alone does not unlock it. Each participant and the synthesis are billed separately against the same organization and API key spend cap.

​
Request

Choose available logical model IDs from the live catalog. models and messages are required. Optional fields include temperature, top_p, max_tokens, and user. Hosted web_search and web_fetch options apply to every participant; function tools are not supported. See the endpoint schema for the full request shape. An optional UUID Idempotency-Key produces stable billing keys for each participant and the synthesis. It is not a cache of the complete response; do not assume a repeated request avoids inference work.

​
Response

Illustrative successful response:
  • complete: at least two participants and the synthesis succeeded.
  • partial: only one participant succeeded, or the synthesis failed.
  • failed: every participant failed. Provider failures can still produce HTTP 200; inspect status and each entry’s error and error_code. Requests where all participants fail deterministic caller validation can return a 4xx error instead.
Participant usage is in prompt_tokens and tokens; synthesis usage is in synthesis_prompt_tokens and synthesis_tokens. There is no aggregate customer_cost field.

​
Privacy rules

Anonymous and Private models can mix. Confidential models only combine with other Confidential models, with a Confidential synthesis and attestation fields in the result. Encrypted participants are rejected by this server route. The Browser App coordinates Encrypted Fusion in the browser with an explicit synthesis step.

​
Through MCP

Agent hosts can use nito.chat_fusion on the hosted MCP endpoint with an OAuth client holding mcp:inference. That tool takes a request ID, prompt, and two or three logical model IDs. Its web search and fetch are on by default with per-call opt-outs.