Architecture

crew/
├── apps/
│   ├── server/        Hono: REST + WebSocket, serves the web UI
│   └── web/           React 19 + Vite + Tailwind v4
├── packages/
│   ├── core/          Agent runtime, LLM providers, tools, prompt, worker
│   ├── memory/        SQLite store: agents, messages, memories (FTS5 + vectors), tasks
│   └── channels/      Telegram (grammY)
├── cli/               `crew`, published as @vefimov/crew
├── docs/              This site (VitePress)
└── docker/

Data flow

 Web UI ─┐
 CLI ────┼──▶ Crew runtime ──▶ Provider (Ollama / OpenAI-compatible)
 Telegram┘        │  ▲
                  ▼  │ tools: remember · recall · list/read/write files
               SQLite (one file)
                  │
                  ▼
           Task queue ──▶ memory extraction, embedding backfill

Design choices

Crew has no native dependencies. node:sqlite ships with Node 22, so npx works without compiling anything.

All state sits in one file, ~/.crew/crew.db, in WAL mode. That covers agents, chats, memories, vectors, settings and the job queue.

Memory search uses FTS5 plus a brute-force cosine scan instead of a vector database. That is plenty for thousands of memories per agent and needs no extra service.

Providers sit behind a small Provider interface (chat, embed, models, ping) with streaming and tool calls. The Ollama and OpenAI-compatible adapters are one short file each.

The runtime is an EventEmitter. The server forwards its events to WebSocket clients, so every UI sees every agent live.

Each agent handles one turn at a time, whichever channel the message came from.