SenteLabsAI has published OpenExecutive, an open-source application that presents a single virtual executive adviser while routing work to eight specialist AI agents behind the scenes. The project combines a FastAPI backend with a Next.js interface and is designed to tailor its responses to an organisation's own profile and documents.
The software keeps the internal multi-agent structure out of the user-facing conversation. Instead, answers arrive through one consistent executive persona. Two retrieval layers support specialist calls: a built-in collection of business material stored in ChromaDB, and a separate collection made from documents uploaded by the user. The retrieved context is added to the current user turn rather than to the cached system prompt.
OpenExecutive also records selected context between sessions. After each response, a background Claude Haiku 4.5 process extracts decisions, initiatives and advice into SQLite. At the start of a later session, the application supplies those past decisions to the executive agent. A scheduler can surface planned follow-ups and time-sensitive actions without waiting for a fresh question.
That scheduler imposes an operational limit. Its database-claiming method is designed for a single API instance, and the maintainers warn that horizontally scaling the API without an additional gate could trigger an action more than once. The supplied Fly.io configuration therefore caps the service at one running machine.
The application requires Python 3.11 or later and Node 22 or later. Its initial setup downloads machine-learning dependencies and an embedding model of about 90 MB to create the local vector index. The API serves on port 8000 and the interface on port 3000. An onboarding flow collects a company profile, after which users can add items such as pitch decks, financial models or strategy documents for retrieval.
Although the default configuration can use Anthropic models, the provider layer also supports OpenAI-compatible local servers including Ollama, LM Studio, vLLM and llama.cpp, as well as routing through OpenRouter. Individual specialists can be assigned different available models. The maintainers note that local configurations do not provide equivalents for Anthropic prompt caching, extended thinking or the project's server-side web-search feature.
The repository includes Discord access through direct messages, mentions and slash commands. It also describes separate development and quality-assurance deployments, Google sign-in with an email allow-list, and a shared-secret boundary between the web interface proxy and backend.
For testing, the project contains 29 scenarios across all eight specialist areas. An automated judge scores persona consistency, domain accuracy, use of company context, routing and actionability on a five-point scale. The stated continuous-integration threshold is an average of at least 3.5, with failures also triggered by a decline of more than 10% in any dimension compared with the main branch.
