Skip to content

Cursor

Cursor supports both ACP (full agent) and MCP (tool-only) integration with AgentZero.

Recommended: ACP — AgentZero runs the full agentic loop (inference, tool calling, approval, context management) using your local models. Cursor acts as the UI.

ACP runs AgentZero as a complete coding agent. Cursor sends messages; AgentZero drives the model, executes tools, and streams results back.

Terminal window
az init --editor cursor

This creates .cursor/rules with ACP integration instructions.

In a Cursor terminal panel, run:

Terminal window
az serve

The server communicates over stdio (JSON-RPC). No network ports are opened.

Send messages to the ACP server via Cursor’s terminal. AgentZero handles model inference and tool execution locally.

Use MCP when you want Cursor’s built-in AI to drive inference, but use AgentZero’s policy-enforced tools (read, write, search, edit, shell).

Open Cursor Settings > MCP Servers, and add:

{
"mcpServers": {
"agentzero": {
"command": "az",
"args": ["mcp"]
}
}
}
ToolDescription
read_fileRead file contents
list_directoryList directory contents
search_filesSearch for patterns in files
write_fileWrite content to a file (requires approval)
run_commandExecute a shell command (requires approval)

All tools enforce AgentZero’s policy engine — path validation, sensitive file blocking, and audit logging apply.

Use CaseProtocol
Full local coding agent with tool callingACP
Cursor’s AI + AgentZero’s secure toolsMCP
Working with sensitive/PII data locallyACP (data never leaves your machine)
  • Stdio-only transport — AgentZero never opens a network port for editor communication
  • Policy enforcement — all tool calls go through the policy engine
  • Audit trail — every action is logged to .agentzero/audit/
  • “command not found: az” — ensure az is in your PATH. Run which az to verify.
  • MCP tools not showing — rebuild with cargo build --features mcp
  • Provider errors — run az doctor to check provider connectivity