OpenExecutive

Open Executive is a virtual executive team for a company's leaders. A person talks to one executive persona, and behind it an orchestrator routes each question to specialist agents in parallel and returns a single synthesized answer. The specialists cover strategy, finance, people, legal, operations, marketing, product, sales, and board communications, and the internal agent structure is never shown to the user.

Each specialist retrieves context from two layers stored in Chroma: built-in business knowledge written as Markdown, and the company's own uploaded documents. A first-run wizard captures a company profile (business model, competitors, priorities, culture), and every answer draws on it. After each response a background pass extracts decisions, initiatives, and advice into SQLite as episodic memory, so later sessions open with what the executive recommended before. A built-in scheduler surfaces follow-ups and time-sensitive actions, and an optional per-person notes feature keeps private, dated records of what each person said.

The executive meets people in Slack, Discord, Telegram, Google Chat, email, and a web UI, and reads and sends mail and books calendar time through Google Workspace or Microsoft 365. With Act as me switched on, it drafts replies in a leader's own Gmail for them to review and send. Google Drive, OneDrive, Confluence, and Notion content syncs into the knowledge base, labeled as unreviewed so it ranks below uploaded documents. The executive also consumes tools from other MCP servers, with deny rules for destructive ones, and exposes its company context and specialist council as an MCP server for clients such as Claude Code and Cursor.

Models run on the Anthropic Claude API by default, with Claude Opus on the deep-reasoning specialists and each agent's model selectable in a council UI. Any OpenAI-compatible server (Ollama, LM Studio, vLLM, or llama.cpp) or hosted gateway can serve some or all agents, though web search, prompt caching, and extended thinking are disabled for those models. An evaluation suite of 29 scenarios is scored by an LLM judge on persona coherence, domain accuracy, use of company context, routing, and actionability, and gates changes in CI.

The system deploys as a FastAPI backend and a Next.js UI with one persistent volume, behind Google or OpenID Connect sign-in with an email allow-list. The API runs as a single instance because the scheduler is not safe across replicas.