Custom providers (BYOK)

Custom Provider lets you connect any API that follows the OpenAI or Anthropic format. This includes large cloud services (OpenRouter, Groq, Mistral, and others) as well as local proxies and self-hosted servers.

Adding a provider

Go to Kodik settings → Provider → Custom Provider → Add provider. Fill in the following fields:

Field	Description
Name	A display name of your choice
Base URL	The root address of the API (see URL requirements)
API Key	Authorization key (optional if the endpoint does not require one)
Format	openai (default) or anthropic
Custom headers	Additional HTTP headers as key-value pairs
Streaming	Enabled by default; disable if the provider does not support it

After you click Continue, Kodik checks the endpoint, fetches its model catalog, and asks you to pick the models you want — at least one is required. The provider is created only when you confirm your selection; closing the wizard without confirming adds nothing.

Subscription requirement

Adding, editing, or switching to custom providers requires a signed-in paid Kodik subscription. Signed-out users, free-plan users, and sessions whose sign-in token cannot prove the subscription see Add provider locked, and backend settings updates reject custom-provider changes. Removing an existing provider is still allowed.

URL requirements

• Remote (non-local) endpoints must use https://. Using http:// is only allowed for localhost, 127.0.0.1, ::1, .local addresses, and private IP ranges (10.x.x.x, 172.16–31.x.x, 192.168.x.x).
• /responses endpoints are not supported — use a base URL or a /chat/completions path.
• If you enter a URL ending in /chat/completions (OpenAI format) or /messages (Anthropic format), Kodik automatically strips the suffix and infers the correct base path.

Format and URL mismatch

Kodik will show an error if the URL ends in /messages (Anthropic) but the selected format is openai, or vice versa. Make sure the format matches your provider’s API.

Model discovery

When you add or update a provider, Kodik fetches the model list from the /models or /v1/models endpoint. Discovered models appear in the model selector along with metadata: name, description, context window size, and image/tool support flags if the provider returns them.

Each model row in the provider card shows VISION, TOOL USE, and THINKING capability badges — they are clickable toggles (colored = on, dimmed = off). If automatic detection doesn’t match what your model actually supports, click a badge to flip that capability.

The “Context” value on a model row is also click-to-edit. Set the real window for models whose provider doesn’t report one — Kodik otherwise assumes 128k, which makes small local models overflow mid-task. A manually set value sticks across catalog refreshes; clear the field to return to automatic.

LM Studio

Point the provider’s base URL at http://localhost:1234/api/v0 — LM Studio’s native REST catalog reports each model’s capabilities and the context length it is actually loaded with, so badges and context fill in automatically.

Reasoning-capable models get a per-model reasoning selector in the chat model picker, the same control Kodik models have. When the provider’s catalog advertises the model’s reasoning modes (LM Studio’s native API reports the exact options and the default), Kodik offers exactly those, with the provider default preselected. Models whose catalog advertises nothing get the standard effort list (None…X-High) with Medium preselected. Your choice is saved on the model and applied to every request — as reasoning_effort for OpenAI-format providers, as OpenRouter’s unified reasoning object for OpenRouter, or mapped to a thinking budget for Anthropic-format providers.

If the first model-catalog response is empty, Kodik waits briefly and retries before showing the picker. If the models endpoint is unreachable or still returns an empty list, Kodik still lets you use the provider — just select the model name manually.

Health indicator

A colored dot is shown next to each provider in the list:

Color	Status	Meaning
Green	ok	Provider responds quickly
Yellow	degraded	Response took more than 3 seconds
Red	unreachable	Provider is down or returned an error

Kodik checks each provider’s health every minute by fetching its model list. The status updates automatically without a page reload.

Known provider auto-detection

Kodik recognizes popular services from the base URL and automatically applies appropriate defaults:

Service	URL	Default format
OpenRouter	openrouter.ai	openai
Groq	groq.com	openai
Together AI	together.xyz	openai
Mistral	mistral.ai	openai
Perplexity	perplexity.ai	openai
Fireworks	fireworks.ai	openai
DeepSeek	deepseek.com	openai
LM Studio	localhost:1234	openai
LiteLLM	localhost:4000	openai

For unrecognized URLs Kodik defaults to openai format. You can override any setting manually.

Detection only affects branding and connection defaults — it never limits what a provider can do. Every model starts with tool use enabled regardless of where the provider is hosted, so agent features (file edits, terminal, MCP servers) work out of the box; a capability is only turned off when the provider’s own model catalog reports that the model does not support it.

Custom headers

The Custom headers section lets you add arbitrary HTTP headers to every request. This is useful, for example, for authentication via X-Custom-Auth or for passing provider-specific routing parameters.

Editing and deleting providers

Click a provider card to open its settings and make changes. Delete a provider using the Delete provider button inside that card.

Caution

If you delete a provider that is selected as the sub-agent provider, its sub-agent settings are automatically cleared. See sub-agent model.