How it works

The office engine (src/lib/server/engine/) starts with the server and checks the queue every second. Each task is given to the employee's model together with the client's request, the manager's plan, what colleagues finished, and the project's files. The model then works by calling tools until it replies with a summary:

Tool	Effect
name_project	Names the folder of a new project
list_files, read_file	Look at the project as it is now
write_file, edit_file, delete_file	Create, change or remove a file in the project folder
delegate	Hands a sub-task to a more junior colleague
ask_manager	Asks the manager who assigned the task; the answer comes back at once
ask_client	Pauses the task and asks you
save_note	Writes down a lesson for future projects
web_search, fetch_url	Look something up (Claude uses Anthropic's own web tools instead)
post_to_slack	Tells you something in Slack, when it is connected
run_checks	Checks syntax and runs tests, when you turned that on

Every task that changes files becomes one git commit in the employee's name. A model that cannot call tools (for example llama3 on Ollama) is given the same abilities as tags in plain text, such as [FILE: path] … [/FILE]; it has no web access and cannot run checks.

The guardrail limits, the per-model prices and the discounts for cached input used for cost estimates live in src/lib/officeConfig.ts. Costs are estimates from list prices; local Ollama models count as free.

Technology

Part	Technology
Frontend	Next.js (React), framer-motion
Backend	Next.js API routes, plus an in-process engine that runs the tasks
Database	better-sqlite3
AI	OpenAI SDK (OpenAI and Grok), Anthropic SDK, Gemini and Ollama HTTP APIs, all streamed, with native tool calling
History	isomorphic-git (no git program needed)

How the server, the engine, the providers and the project folders fit together is described in ARCHITECTURE.md. Where each part lives in the code is described in CONTRIBUTING.md.