Run an open-source Claude Cowork alternative on your own model or a local LLM

What "local LLM" means when an agent has to run

A local LLM is a language model that a team runs on infrastructure it controls, or reaches through an endpoint and key it holds, instead of a model a vendor serves only inside its own product. The distinction matters most for an agent, because an agent makes many model calls to finish one job. The model choice decides which provider answers each step, who holds the API key, and where the work runs. Kortix is the open-source AI Operating System, and it treats that choice as project configuration rather than a fixed menu.

Kortix runs any provider, on keys you hold

Kortix is model-agnostic. Kortix states the policy in these words: "Any model provider with your own keys. Or the ChatGPT plan you already pay for. Or sign in with your OpenCode Console account for OpenCode Zen and Go." You pick the model per agent, per session, or per message, and you can switch the day a better model lands. The same project can mix a model Kortix manages, a model behind a key you hold, and a subscription you already pay for. On Kortix Cloud, Kortix sends each managed model only to a fixed pool of US inference endpoints with zero data retention.

The deployment path: connect, set a default, pin per agent

Every model request goes through the LLM gateway, which finds the credentials, applies the project’s routing and access rules, and records the cost. A model id tells the gateway where the credentials come from. A bare managed id such as kimi-k3 is billed in Kortix credits. A provider/model id such as anthropic/<model-id> runs on your provider API key, and your provider bills you. A codex/<id> id runs on your ChatGPT Plus or Pro subscription. Kortix charges nothing for a request that runs on your own key or subscription. You connect a provider with one command, such as kortix providers set anthropic, set a project default with kortix models default <model-id>, and pin one agent with kortix agents model <agent> <model-id>. A session that names no model follows that chain, from an explicit pin down to the platform default. Read the docs for the full model configuration.

Self-host the stack and point it at your own endpoint

A team that has to keep the machine in its own network can run the whole platform itself. Self-host is free, and a self-hosted instance has no Kortix-managed models, so you connect your own providers. A self-hosted or unlisted OpenAI-compatible endpoint connects through the Custom tab, where you enter a provider id, a base URL, a model id, and a model name; the key is optional. That endpoint calls your server directly and is read by the OpenCode harness. The self-hosting walkthrough covers the install, and the self-hosted AI agent platform page covers what a team runs once the stack is up.

What a closed desktop assistant does not give you here

Claude Cowork is Anthropic’s closed desktop assistant, and ChatGPT Work is OpenAI’s. Each runs on one vendor’s models: Claude Cowork on Anthropic models only, ChatGPT Work on OpenAI models only. Neither runs on a model you host or a key you hold, and neither exposes its configuration as files in a repo you own. Kortix is the open-source alternative for the case where the model, the keys, and the runtime have to be yours. The source is public: read it, fork it, audit it. Kortix on GitHub carries the code, and the README’s comparison table lists what each product allows as of July 2026.