// the find
SenteLabsAI/OpenExecutive
AI-powered virtual executive team — a single coherent executive persona backed by 8 specialist agents (FastAPI + Next.js).
An opinionated multi-agent LLM app that fronts nine specialist Claude-backed agents (CFO, GC, COO, etc.) with a single 'executive' persona, chatable over Slack, Discord, Telegram, email, or a Next.js UI, with RAG over your own company docs and an MCP server so other agents can query it. Aimed at founders/small teams who want a Claude-powered advisory layer without building the orchestration themselves.
The prompt caching design is genuinely careful — persona/company profile/knowledge index cached separately from RAG context which is injected into the user turn, and they claim 85% cache hit rate after a few turns, which is the right way to keep multi-agent cost down instead of re-sending the whole system prompt every call. The eval harness isn't decorative: 29 scenarios scored by an LLM judge across 5 rubric dimensions, with a CI gate that fails a PR if any dimension regresses >10% vs main, so prompt changes actually get checked instead of eyeballed. Provider abstraction is real, not just a config flag — you can run fully on local Ollama/vLLM models or an OpenRouter gateway with no orchestrator code changes, and the docs are upfront about what breaks (no caching, no extended thinking, weaker tool-use routing) rather than hiding the tradeoff.
The scheduler claims jobs via `UPDATE ... RETURNING`, which only works single-instance — there's no distributed lock or queue, so this can't horizontally scale without someone rewriting that piece first, and it's a hard ceiling baked into the architecture rather than a config toggle. Vector storage is embedded ChromaDB plus SQLite for episodic memory, fine for one company's data on one box but with nothing described for multi-tenant isolation if the hosted 'openexecutive.ai' offering needs it. The 'Act as me' feature drafts and sends email from someone's real Gmail in their voice, and the system handles spend approval thresholds — both are meaningful blast-radius features and the README leans on 'review before send' as the entire safety story, with no detail on what stops a bad specialist output from going out under a real signature. Running both the embedded Discord bot and the standalone `make discord` process against the same token double-answers every message — a footgun that's documented but still easy to hit during dev.