Open Executive, an open-source system that assigns executive roles such as CFO, COO, and general counsel to eight AI agents while presenting a single executive voice to the user, is drawing attention. It was built by Sente Labs, a firm that works on multi-agent systems, and released under Apache 2.0. The design assumes you run it on your own infrastructure and feed it your own company documents.

Eight Specialists, One Executive Voice

The notable choice in Open Executive is that the internal structure is never exposed to the user. Questions go to a component called the Executive Orchestrator, which calls the relevant specialists in parallel and merges their views into a single answer. From the outside, it always looks like the same executive persona speaking.

Responsibilities are split across 8 domains. The chief strategy officer handles competitive analysis, M&A, and objectives. The chief financial officer covers financial modeling, fundraising, and unit economics. The chief HR officer takes hiring, compensation, and culture, while the general counsel handles contracts, intellectual property, and compliance. The chief operating officer works on process design and vendor management, the chief marketing officer on go-to-market strategy and brand, the chief product officer on roadmap and prioritization, and a board communications director on board decks and investor relations.

Model selection follows the same split. The default is claude-sonnet-4-6, but strategy, finance, legal, and board work call claude-opus-4-7 with extended thinking. Expensive models are reserved for the domains where judgment carries the most weight.

Memory That Makes "You Said That Last Month" Possible

The feature that separates this from a chatbot is memory that survives across sessions. After each response, a background claude-haiku-4-5 pass extracts decisions, ongoing initiatives, and advice from the conversation and writes them to SQLite. The next session opens with a block of past decisions, so the discussion resumes where it left off.

Knowledge sits in 2 layers. The first is built-in MBA-level material written in Markdown and loaded into the ChromaDB vector store at startup. The second is your own uploaded documents, whether a pitch deck, a financial model, or a strategy memo, which are chunked into a separate collection and retrieved only when a question calls for them. Retrieved context is injected into the user turn rather than the system prompt, specifically to avoid breaking the cache.

That cache is deliberately structured too. The executive persona, the company profile, and the knowledge index are cached separately, reaching a cache hit rate of up to 85 percent after the first few turns. The longer the conversation runs, the more that matters for cost.

Running Without Claude Is an Option

Anthropic models look like a hard requirement, but the system runs without an API key. Point it at any OpenAI-compatible local server such as Ollama, LM Studio, vLLM, or llama.cpp and everything from the orchestrator to the specialists runs on your own hardware. Routing calls through OpenRouter is also supported, so you can keep the executive on Claude while switching individual specialists to a local model.

The limits are stated plainly. Server-side web search, prompt caching, and extended thinking are unavailable on local models. The orchestrator also picks specialists through tool use, so a small model that is weak at tool calling will route poorly. Getting it running is easy, but answer quality tracks model choice directly.

Grading Executive Advice in CI

For anyone wondering how advice quality is kept in check, there is an evaluation suite. It contains 29 scenarios covering all 8 domains, each defining a simulated company situation, the topics that should come up, the specialist that should be invoked, and a scoring rubric. The judge is claude-opus-4-7, which rates persona coherence, domain accuracy, use of company context, routing quality, and actionability on a 1 to 5 scale.

An average below 3.5 fails CI, and any pull request that drops a single dimension by more than 10 percent against main is rejected. Adding a new specialist agent requires at least 2 new evaluation scenarios. Keeping a model-judged regression test on something as subjective as executive advice is unusual for a project of this kind.

Setup Effort and What You Hand Over

The stack is FastAPI on Python 3.11 or later plus Next.js 15. Clone the repository, set one environment variable, and run a make command. The first launch takes a few minutes because it downloads an embedding model, but later starts are fast.

git clone https://github.com/SenteLabsAI/OpenExecutive.git
cd OpenExecutive
cp .env.example .env
make dev

The web UI is not the only entry point. Slack, email, Telegram, Google Chat, Discord, and the command line all reach the same executive. On Discord it responds to mentions as well as slash commands.

As for handing over company material, the company profile, uploaded documents, and the vector store are all excluded from version control and stay on your machine or on a cloud volume you control. The only thing that leaves is the portion of the prompt sent to the Anthropic API. Still, this is a system built around reading your financial model and board decks in full, so where to draw the line remains an operational decision.

Summary

Open Executive is less about handing management over to AI and more about keeping an always-available advisor that knows your company context. The repository has 3,246 stars and 306 forks as of this update, and Apache 2.0 permits commercial use. What stands out is not the 8-agent structure itself but the work put into memory and evaluation to make the output reproducible.