Providers

Set up a custom provider

Connect an endpoint that implements OpenAI Responses, OpenAI Chat Completions, or Anthropic Messages with one of Kition’s supported authentication schemes.

Updated

By Kition Docs

Verified against Kition 0.2.0 (7d49036f) on 2026-11-06

Compatibility boundary

A custom provider must implement one of the three wire APIs Kition supports: OpenAI Responses (responses), OpenAI Chat Completions (chat_completions), or Anthropic Messages (anthropic_messages). Choose the one the endpoint actually implements.

Model discovery is separate from generation. A successful models request proves that Kition can read a model list; it does not prove that streaming, tool calls, or Agent generation work on the selected wire.

Heads up

Hosted web search is available only on the Responses and Messages wires, not on Chat Completions. Verify the exact wire before entering a production credential.

Fields available in Settings

  • Base URL — the root used for the provider API
  • API key — the token Kition sends with provider requests
  • Models endpoint — /models by default
  • Auth scheme — Bearer, Raw, or X-API-Key
  • Wire API — OpenAI Responses, OpenAI Chat Completions, or Anthropic Messages
  • Reasoning effort — Minimal, Low, Medium, or High; Extra high and Max also appear for Anthropic Messages

Configure the endpoint

A bare host such as https://llm.example.com is saved with /v1 appended; an explicit path is kept as entered. Enter the complete documented Base URL when a gateway uses a nonstandard prefix.

  • Open Settings → AI Providers and select Custom
  • Enter the exact Base URL supplied by the endpoint operator
  • Enter a limited API key or token
  • Open Advanced and set the models endpoint, auth scheme, and wire API
  • Choose Sync now, then select a returned text model in the Agent model picker
  • Send one small, non-sensitive prompt before trusting the endpoint with workspace context

Headers and authentication limits

Settings lets you choose the auth scheme, but it does not expose an editor for arbitrary extra headers or a custom header name. Endpoints that require tenant headers, beta headers, signed requests, or several independent credentials may need a gateway that presents a supported interface.

Do not place additional secrets in the Base URL, model name, or workspace files as a workaround.

Local and private endpoints

Kition accepts the Base URL you enter, including a local HTTP service. Use plain HTTP only on a trusted local connection; use TLS for traffic that leaves the device.

A self-hosted endpoint can receive prompts, selected workspace context, tool definitions, and generated outputs. Review its logs, retention, and access control before connecting sensitive workspaces.

Diagnose in layers

  • Sync fails — check Base URL, models endpoint, auth scheme, DNS, proxy, TLS, and service availability
  • Models sync but the Agent fails — verify the wire API, streaming format, selected model, and tool-call support
  • 401 or 403 — verify token format and permissions
  • 404 — confirm both the models endpoint and the generation endpoint for the chosen wire

Credential storage

Custom-provider credentials use the same local plaintext secret-file storage as built-in providers. Disconnect disables the provider and deletes that saved secret; revoke the remote credential when it may be exposed.

Implementation sources

Related pages

Ready when you are.

Kition is a desktop AI workspace for connected documents, structured tables, editable designs, whiteboards, visual workflows, and reviewable AI agents.