Models

Every workspace picks its own model, and you can change it whenever you like. There are two ways to connect one.

The model does not decide what an agent can reach

Whatever is behind an agent — a frontier model or a small one on your own machine — it is given the same tools: web search and fetching, reading and writing your knowledge base, searching it, running scripts, and every account you have connected.

Rookery supplies those itself rather than relying on the model to have them, so the model determines how well a job is done, not whether it can be done at all. Tools remove the floor, not the ceiling. This is why a modest local model is a reasonable choice for routine agents.

Where that stops being true

Browser control is the exception. A connected account’s actions are curated and pre-authenticated — a narrow, well-defined surface, which is exactly what lets a small model use them reliably. Driving a live web page is the opposite: open-ended, stateful, and unforgiving of a wrong click.

So an agent that reads and acts through connections works well on a modest model, while one that signs in and clicks through a site still rewards a capable one. Worth knowing before you choose a model for an agent that does the latter.

A coder tool you already have

If you already use a command-line coding assistant, Rookery can drive it directly and reuse the sign-in you already did. No key to paste.

Supported: Claude Code, Codex, OpenCode, Cursor and Gemini CLI. Rookery looks for them on your PATH and in ~/.local/bin.

Each workspace gets its own isolated configuration directory, so one workspace’s agents can never read another’s session or history. Your operator credentials are copied in per invocation.

Caution

OpenCode needs an explicit model. Unlike Claude Code, whose model comes with its login, OpenCode talks to many providers and has no default of its own. With none set it targets a hardcoded provider and fails with an authentication error — which looks like a broken login but is really a missing model.

Because each workspace is isolated, OpenCode does not inherit your own opencode.json, so setting a default there does not reach Rookery. The model has to be set on the workspace.

A provider API

Give Rookery a provider, a model name and a key, and it talks to the API directly in-process — no separate tool involved.

Hosted covers the frontier labs (OpenAI, Anthropic, Gemini, xAI, Mistral, DeepSeek, Moonshot, Z.AI), the routers (OpenRouter, Perplexity, OpenCode Zen and Go), the enterprise clouds (AWS Bedrock, Alibaba Cloud) and the open-weight inference clouds (Groq, Together, Fireworks, Cerebras, SambaNova, Nebius, DeepInfra, Hugging Face, GitHub Models, Ollama Cloud).

Self-hosted covers OpenAI-compatible servers on your own hardware: Ollama, vLLM, LM Studio, llama.cpp, LocalAI and Jan. These need no key — point Rookery at the address and nothing leaves your network.

There is also a Custom (OpenAI-compatible) option for anything else that speaks the same protocol.

Note

Azure OpenAI and Google Vertex AI are not supported. Azure puts the deployment name in the path and requires a version parameter; Vertex mints short-lived tokens from a service account. Neither fits the “provider, model, key” shape, and each needs its own implementation rather than a catalogue entry.

Setting it up

Settings → Coder in the workspace. Pick the kind, then the provider and model. You can paste a key straight into the form and Rookery stores it as an encrypted secret for you.

The base URL is prefilled with the provider’s default and can be overridden — which is how you point at a local server on a non-standard port. An untouched prefill keeps following the default rather than freezing on today’s address.

The container

The published image ships no coder tool and enforces it: a workspace running in the container must use a provider API. Attempting to configure a local coder fails with a message saying so, rather than trying to run a binary that is not there.

Which to choose

If you	Use
Already have a coder tool set up	That tool
Want the fewest moving parts	A provider API
Want nothing leaving your network	A self-hosted model
Run the container	A provider API

Cost and limits

When a provider runs out of credit or rate-limits you, the run fails with a message saying which — “out of quota” and “try again shortly” are different problems and are reported differently. Token usage is recorded per run on the agent’s page.