agent acp

Starts the CLI in ACP server mode (Agent Client Protocol) for editor and client integrations.

agent acp starts Cursor Agent as an ACP server over stdio, using newline-delimited JSON-RPC 2.0 messages, e.g. for editor integrations like Zed or JetBrains. Per the docs the command is 'hidden' and meant for building custom clients, not normal terminal use. Authentication happens beforehand via agent login, --api-key/CURSOR_API_KEY, or --auth-token/CURSOR_AUTH_TOKEN; endpoint and TLS options (-e, -k) are appended before acp, e.g. agent -e https://api2.cursor.sh acp.

✅ WHEN TO USE IT?

Build a custom client or editor integration against Cursor Agent

agent acp — then send initialize/session/new/session/prompt via JSON-RPC over stdin/stdout

Authenticate up front with an API key instead of logging in through the ACP flow

agent --api-key "$CURSOR_API_KEY" acp

⛔ WHEN NOT TO?

Normal terminal use without building a client

agent acp only outputs raw JSON-RPC messages on stdout, not a readable chat UI.

Better: Just use agent (interactive mode) or agent -p for scripts

External tools/data need to be made available to the agent

ACP is for client integrations, not for wiring tools into the agent.

Better: Use agent mcp to connect MCP servers for tool access

SOURCES

Cursor Docs — CLI Reference: Parameters ↗ cursor.com
Cursor Docs — CLI ACP ↗ cursor.com