Skip to content

Your models & keys (BYOK)

You can connect your own API keys and use your own models alongside the platform ones. Open Account → Models. Bringing your own keys may not be included on your plan — see donkit.ai/pricing.

Providers

Add a provider with your credentials — every provider the platform itself supports is available:

  • Anthropic — your Anthropic API key.
  • OpenAI-compatible — your OpenAI key, or any OpenAI-compatible endpoint (OpenRouter, vLLM, a gateway) via a custom base URL.
  • Azure OpenAI — endpoint + key.
  • Vertex AI (Gemini), Claude on Vertex, Vertex OpenAI — a Google Cloud service-account JSON.
  • Claude on Azure Foundry — Foundry endpoint + key.
  • Ollama — a publicly reachable Ollama URL.

Credentials are encrypted at rest and never shown again after saving — the page displays a masked preview. Use Test to verify a key (not available for Vertex-based providers). Custom endpoints must be reachable from the internet.

An account can hold up to 50 providers, 200 chat models and 50 embedding models. If you reach a limit, delete something you no longer use.

Chat models

Add a model served by one of your providers: give it a name, the model ID your provider expects (for example gpt-4o or claude-sonnet-4-5), its model family, its context window, and which capabilities it supports (tool calling, vision, files and so on). Name it whatever you like — we add a BYOK label to the name, so it never collides with a platform model. The model appears in every model picker for your account — agents, sub-agents and presets can use it.

Model family tells the platform which transport rules to apply — how to control the model's thinking, which reasoning fields to expect, how to parse tool calls. Pick the family the model actually belongs to (Claude, GPT‑5, GLM, Kimi, DeepSeek, …), or Generic OpenAI-compatible for anything else. The form suggests a family from the model ID; check the suggestion before saving, because a wrong family means the model receives the wrong request shape and may answer with errors or stop calling tools. For an OpenAI-compatible provider with a custom base URL you also pick the hosting (OpenRouter, Fireworks, vLLM, …); every other provider type fixes it for you.

For GLM on a self-hosted server (vLLM, z.ai or a generic OpenAI-compatible endpoint) the thinking depth maps onto the model's own scale: Off disables thinking, Low sends the reduced level, and every other builder setting sends the model's full budget. On GLM‑5.3 the template's high level is a reduced one, and a builder running on it stops calling tools — so High in the platform never becomes high on the wire for these hosts.

Usage of your own models is not charged platform credits — you pay your provider directly. Usage is still recorded in your account usage history.

Auto model selection always uses platform models. Your own models are picked explicitly.

Embedding models

You can also add an embedding model (OpenAI API only) and make it the default for your account — document ingestion and retrieval will then use it. Set the correct vector dimensions (e.g. 1536). Changing the default embedding model requires re-ingesting existing knowledge bases: collections indexed with a different model stop matching and ingestion will report a dimension error.

Turning platform models off

The Platform models toggle removes platform models (including Auto) from the pickers for your account, leaving only your own models selectable.

The switch applies to running agents too, not just to the pickers: an agent still configured with a platform model (or with Auto) stops answering and reports that platform models are turned off, until you point it at one of your own models. That is deliberate — otherwise an agent configured before you flipped the switch would keep spending platform credits. The builder chat behaves the same way: with the toggle off it refuses Auto and asks you to pick one of your own models.

If your plan later loses the Your models capability, the toggle stops applying: your own models become unavailable and the platform catalog comes back, so you are never left with nothing to run.