Providers & models

Each employee runs on one provider and one model. You can mix them freely: a strong model for the people who plan and review, a cheaper one for those who carry out well-defined work.

Providers

Provider	Key	Get a key	Default model
OpenAI	OPENAI_API_KEY	platform.openai.com	gpt-4o-mini
Anthropic (Claude)	ANTHROPIC_API_KEY	platform.claude.com	claude-opus-5-5
Google Gemini	GEMINI_API_KEY	aistudio.google.com	gemini-2.5-flash
xAI (Grok)	GROK_API_KEY	console.x.ai	grok-3
Ollama (local)	none	ollama.com	llama3

Enter keys in the app's Settings (the gear button), or in .env.local when running from the source. A key saved in the app takes precedence over the environment.

A provider without a key answers with mock text instead of calling a model. It costs nothing and lets you try the interface, but nothing is built.

Choosing a model

When you hire someone, pick the provider and type or choose a model. The hiring form suggests a few per provider; any model name the provider accepts works. Leave the model empty to use the provider's default, which you can change in Settings (or with OPENAI_MODEL, ANTHROPIC_MODEL, GEMINI_MODEL, GROK_MODEL, OLLAMA_MODEL).

To change a model later, edit the employee: their history stays.

A starting point

• CEO and Tech Lead — they plan, delegate and do the final review, so their judgement matters most: give them your strongest model.
• Designer, QA — a mid-range model.
• Junior Dev, Copywriter — well-scoped work: a fast, cheap model is usually enough.

Not sure? Run the same request on two or three models with Insights → Compare models and look at result, time and cost side by side. See Costs & budgets.

If you use a model the office has no price for, its cost shows as unknown; see adding a price.

Tools, and models without them

Employees work by calling tools: listing and reading files, writing and editing them, delegating, asking questions, searching the web (the full list).

Some models — many small local ones — cannot call tools. The office notices when a model refuses them, remembers it, and gives that model the same abilities as tags in plain text instead, such as [FILE: path] … [/FILE]. The work still gets done, with two differences: no web access, and no checks.

Web access

With web access on (the default), employees can search the web and read public pages for current facts. How a search runs depends on the employee:

• Claude employees use Anthropic's own web search and fetch tools.
• Everyone else has their searches answered by a search-capable model: their own provider's if it is OpenAI or Gemini, otherwise any provider with a key. The defaults are gpt-4o-mini-search-preview, gemini-2.5-flash and claude-haiku-4-5; change them with OPENAI_SEARCH_MODEL, GEMINI_SEARCH_MODEL and ANTHROPIC_SEARCH_MODEL.

So web search needs an OpenAI, Gemini or Anthropic key somewhere in the office. Searches count towards the task's cost.

Turn web access off in Settings, or with OFFICE_WEB_ACCESS=off. Pages on your computer or private network can never be fetched; see the security model.

Ollama: free and fully local

Ollama runs models on your own computer: no key, no cost, and nothing leaves the machine.

1. Install Ollama and pull a model: ollama pull llama3
2. Hire an employee on Ollama, with that model name (or leave it empty for llama3).

Setting	Default	What it does
OLLAMA_URL (or Settings)	http://127.0.0.1:11434	Where Ollama runs. From Docker, use http://host.docker.internal:11434.
OLLAMA_MODEL (or Settings)	llama3	The model for Ollama employees hired without one.
OLLAMA_NUM_CTX	8192	The context window. The project's files are sent with every task, so raise it for larger projects if your model and memory allow.

Small local models do noticeably weaker work on multi-step tasks than the large hosted ones. A practical mix is a hosted model for the CEO and reviewers, and local models for well-defined work.