crow-cli

A stateful, ACP-native coding agent. No framework — OpenAI SDK, FastMCP, and a shared LanceDB memory store every agent can query.

What we didn't build

Permissions	The agent has the same access you do. You trust it or you don't run it.
New protocol	We use ACP. No new data model for agent-client interaction.
Offline agent	Every agent has web_search and web_fetch and is prompted to use them liberally.
Static skill packages	Skills are markdown the agent loads at runtime — and this site publishes ours.

Start here

• Installation — clone and build from source, like zed
• Configuration — everything in config.yaml
• Zed — running crow from our zed fork (recommended client)
• Services — the daemon layer: memory, MCP, embeddings, searxng
• CLI walkthrough — every part of the crow-cli helper
• Skills — the workflows our agents load at runtime

The shape of it

• crow-cli acp is the agent. Editors (zed, Crow ADE) speak ACP to it over stdio.
• crow-cli run is the client — one-shot prompts, a REPL, and the delegation mechanism between agents.
• crow-memory is a Rust HTTP service over LanceDB. Sessions, messages, and multivector ColBERT embeddings; agents mine each other's history with query_memory / query_session.
• Skills are directories in ~/.agents/skills with a SKILL.md. The system prompt catalogues them; the agent reads one when a task matches.