SenteLabsAI releases Open Executive, an AI virtual executive team
SenteLabsAI, working under the sentelabs.ai brand, has open sourced Open Executive, a system meant to act as a company's virtual executive team. A user talks to a single, consistent executive persona, but behind that voice sit eight specialist AI agents: a Chief Strategy Officer for competitive analysis, M&A and OKRs; a Chief Financial Officer for financial modeling, fundraising and cash flow; a Chief HR/People Officer for hiring, compensation and culture; a General Counsel for contracts, IP and basic employment law; a Chief Operating Officer for process design and vendor management; a Chief Marketing Officer for go-to-market strategy and PR; a Chief Product Officer for roadmap and prioritization; and a Board Communications Director for board decks and investor relations. A user message goes to an Executive Orchestrator running on claude-sonnet-4-6, which calls the relevant specialists in parallel through tool use. Each specialist pulls context from two ChromaDB collections before answering: a git-tracked, built-in library of MBA-level knowledge documents, and a separate collection built from the company's own uploaded documents, such as a pitch deck, financial model or strategy notes. The internal multi-agent structure is never exposed to the user, who only sees the synthesized response.
The system keeps state across sessions in two ways. After every response, a background pass on claude-haiku-4-5 extracts key decisions, initiatives and advice into a SQLite database, so the next session opens with a block of past decisions the Executive can reference. A built-in scheduler separately tracks time-sensitive follow-ups and can surface them proactively; because it claims due jobs with a database UPDATE ... RETURNING statement to avoid double-firing, the documentation states the API must run as a single instance, and warns against horizontally scaling it without first gating the scheduler. On the cost side, the system prompt is structured so the executive persona, company profile and knowledge index are cached separately from the per-turn RAG context, which the project says reaches up to an 85% cache hit rate after the first few turns.
The project is self-hosted: cloning the repository and running make dev starts a FastAPI backend on port 8000 and a Next.js 15 UI on port 3000, requiring Python 3.11+ and Node 22+; the first run also downloads a roughly 90 MB embedding model to build the local vector index, which the documentation says makes the first boot take a few minutes. A first-run onboarding wizard collects the company's basics, business model, competitive landscape, strategic priorities and culture before the Executive starts referencing that context. Deployment configs target Fly.io, with separate dev and QA app sets and an optional Honcho memory app, and integrations exist for Slack, email, Telegram, Google Chat and Discord, the last supporting DMs, channel mentions and /ask and /today slash commands. Beyond the hosted Anthropic API, Open Executive can run against any OpenAI-compatible local server, such as Ollama, LM Studio or vLLM, or route through OpenRouter, letting an operator run without an Anthropic key at all or mix a Claude-based Executive with locally hosted specialists; server-side web search and Anthropic's prompt caching and extended thinking are unavailable in that local-model mode. The repository ships 29 evaluation scenarios covering all eight specialist domains, run through an LLM-as-judge eval suite.
Key facts
- Open Executive merges eight specialist AI agents, covering strategy, finance, HR, legal, operations, marketing, product and board communications, into one consistent executive voice, orchestrated by claude-sonnet-4-6.
- Each specialist call draws on two ChromaDB knowledge layers: a built-in, git-tracked MBA-level knowledge base, and a separate collection built from the company's own uploaded documents.
- A background claude-haiku-4-5 pass writes key decisions and initiatives from every response into SQLite, so later sessions open with a record of past recommendations.
- Prompt caching separates the executive persona, company profile and knowledge index from per-turn dynamic content, reaching up to an 85% cache hit rate after the first few turns.
- The built-in scheduler claims jobs via UPDATE ... RETURNING to prevent double-firing, so the documentation requires the API to run as a single instance and warns against scaling it out without gating the scheduler first.
Why it matters
The HN submission title frames this as developers building an AI CEO after being fired to make room for AI, but the project's own documentation does not narrate any layoff or name a company; it opens directly as a product description. What is verifiable is a working, open-source multi-agent system that packages several patterns other AI-agent projects use individually, parallel specialist tool calls behind one persona, a two-layer RAG setup, cross-session episodic memory and a cache-aware prompt structure, into a single self-hostable product aimed specifically at giving a small company access to executive-level advisory functions without hiring them.
Who it affects
It is aimed at founders and small businesses that cannot afford a full executive team, and at developers evaluating or extending multi-agent orchestration patterns for other domains. The scheduler's single-instance requirement and the choice between the hosted Anthropic API and self-hosted local models also make it relevant to teams weighing data handling and infrastructure cost for this kind of system.
How to use it
The code is on GitHub under SenteLabsAI/OpenExecutive. Cloning it and running make dev starts the FastAPI backend and the Next.js UI locally, requiring Python 3.11+, Node 22+ and, by default, an ANTHROPIC_API_KEY. An onboarding wizard then builds a company profile from basics, business model, competitive landscape, priorities and culture, after which the Executive references that context and any uploaded documents. Operators who prefer not to depend on Anthropic can point it at a local OpenAI-compatible server such as Ollama, LM Studio or vLLM, or route through OpenRouter, and can mix a Claude-based Executive with locally hosted specialists; deployment configs are provided for Fly.io, and chat access can be added through Slack, email, Telegram, Google Chat or Discord.
How solid is it
The repository includes 29 evaluation scenarios covering all eight specialist domains, run through an LLM-as-judge eval suite, and the architecture documentation describes concrete engineering choices, such as separating cached from dynamic prompt content and gating the scheduler behind a single-instance constraint, that suggest real production use rather than a demo. What is not verifiable from the fetched text is any outside account of how well the advice holds up in practice, or the specific incident the HN submission's title alludes to.
Risks and caveats
The scheduler's single-instance requirement means running two API machines would double-fire scheduled actions, a constraint an operator scaling the deployment could miss. Running with the hosted Anthropic API also means company documents uploaded for RAG, financial models, strategy notes and pitch decks, are sent to that API unless the operator switches to local models, which themselves lose server-side web search and prompt caching. The documentation text available here cuts off mid-sentence in the evals section, so what the eval score itself measures is not visible.