Kimchi

Kimchi is a terminal coding agent that also ships a VS Code extension and the Studio desktop app. Its configuration has an llmEndpoint field that defaults to https://llm.kimchi.dev; change it and inference requests go somewhere else.

One caveat up front: the official docs describe llmEndpoint only as the inference address and do not promise that it can point at any third-party OpenAI-compatible gateway. The approach below works in the community, but it is a case of configuring by field semantics, not an officially supported BYOK path, and it may break after an upgrade.

PrerequisitesCopy anchor link

First create a dedicated key for Kimchi in the console (naming it kimchi is recommended), then install the CLI:

brew install getkimchi/tap/kimchi

Or use the install script:

curl -fsSL https://github.com/getkimchi/kimchi/releases/latest/download/install.sh | bash

Configuration to fill inCopy anchor link

Setup stepsCopy anchor link

The config file is at ~/.config/kimchi/config.json, and a project-level .kimchi/config.json can also override it:

{
  "apiKey": "sk-your-API-Key",
  "llmEndpoint": "https://aicentos.com/v1"
}

Custom models are registered in models.json, in the same format as self-hosted providers such as vLLM and LM Studio. The field spec is not fully public, so it is best to get the default model working first and only then consider adding custom entries.

Restart kimchi after the change.

TestCopy anchor link

After starting, send one prompt in the session. If you get a 401, first check that the key has no stray spaces; if you get a 404, check whether llmEndpoint stops at /v1.

Then check usage in the console — this is the key step: if usage does not grow, the requests are not going through llmEndpoint at all and are falling back to the default endpoint, in which case this route does not hold for the current version.

NotesCopy anchor link

• Kimchi requires signing in to its own account; the apiKey and the account sign-in state are two separate things, so do not do only half of it.
• External models are treated as the standard tier with vision: false by default, so multimodal models do not carry images here.
• The default model is picked automatically by its own harness, so you may not be able to name a specific aicentos model directly.
• The official docs do not cover this route, so it works but is not guaranteed; have a fallback ready for production.

Related linksCopy anchor link

• Kimchi CLI docs
• kimchi repository