An Open-Source AI Executive: Eight Agents Instead of a Board

The headline under which the project landed on Hacker News reads like a small act of revenge from the automation era: a CEO fired the developers to make room for AI — and the developers released an open-source AI CEO. Behind the joke sits a thoroughly serious repository, OpenExecutive from the company Sente Labs, published under the Apache 2.0 license.

What it is

The system presents itself as a single “senior advisor” with the expertise of a Harvard MBA graduate, but behind one coherent voice hide eight specialized agents: a chief strategy officer, a chief financial officer, a chief HR officer, a general counsel, a chief operating officer, a chief marketing officer, a chief product officer, and a board-relations specialist. The user talks to a single “executive” — the internal multi-agent kitchen is never shown on the outside.

The architecture works like this: a message reaches an orchestrator running on the claude-sonnet-4-6 model, which calls the necessary specialists in parallel, each of them pulling context from a ChromaDB vector database. There are two sources of knowledge — a built-in “MBA library” in Markdown and the company's own uploaded documents (pitch decks, financial models, strategies). For the “deep” roles such as the financier and the counsel, claude-opus-4-7 with extended reasoning is brought in.

Memory and proactivity

A particular point of pride for the authors is episodic memory. After each response, a background pass on claude-haiku extracts the key decisions and pieces of advice and files them away in SQLite, so that in the next session the “executive” remembers what it recommended a month ago. The built-in scheduler is able to remind you of overdue and urgent tasks on its own — though at the cost of one limitation: the API cannot be run in more than a single instance, otherwise tasks would fire twice.

Technology and privacy

The backend is written in Python and FastAPI, the interface in Next.js 15. You can talk to the “director” through the web, the CLI, and also Slack, Telegram, Discord, email, and Google Chat. Deployment is designed for Fly.io, but the system can be taken off the cloud entirely: it works with any OpenAI-compatible local models (Ollama, LM Studio, vLLM, llama.cpp) or through OpenRouter, and in that case an Anthropic key is not needed at all.

The authors stress that the contents of the company folder — profile, documents, and vector store — never leave the local machine or the user's own volume, except as part of requests to the API, and they remind us that Anthropic does not train on API data.

The project has gathered around two hundred stars and, beyond its technical guts, serves as a curious comment on the debates about automation: it seems that code can replace not only the people who write code.

Source: Hacker News