Agent Client Protocol

Open Interpreter can run as an Agent Client Protocol agent. ACP lets an editor or another UI start an agent process over stdio, create sessions, send prompts, stream assistant messages, show tool progress, and request permissions without scraping a terminal.

Use ACP when you want Open Interpreter inside an editor agent panel or another structured agent UI. Use Non-interactive mode for scripts, the SDK for app-server integrations, and MCP server when another agent needs to call Open Interpreter as a tool.

Start the ACP Agent

Most ACP clients start the agent for you. Configure the client to launch:

interpreter acp

If you are wiring a client directly to the standalone binary layout, you can also launch the ACP binary itself:

interpreter-acp

The process speaks ACP over stdin/stdout. Do not wrap it in an interactive terminal UI.

Environment

The ACP agent uses the same Open Interpreter home, config, auth, provider, MCP, skills, sandbox, and approval settings as the terminal CLI.

Common environment variables:

INTERPRETER_HOME="$HOME/.openinterpreter"
OPENAI_API_KEY="..."
ANTHROPIC_API_KEY="..."
KIMI_API_KEY="..."

If the client lets you pass environment variables, prefer setting provider keys there or in your shell profile. You can also configure providers in $INTERPRETER_HOME/config.toml; see Configuration and Providers.

Zed

Zed has native ACP support in the agent panel. Add Open Interpreter as a custom agent server in Zed settings:

{
  "agent_servers": {
    "Open Interpreter": {
      "type": "custom",
      "command": "interpreter",
      "args": ["acp"],
      "env": {
        "INTERPRETER_HOME": "/Users/you/.openinterpreter"
      }
    }
  }
}

Open the agent panel, create a new external-agent thread, and select Open Interpreter.

JetBrains IDEs

JetBrains AI Assistant supports ACP agents from AI Chat. Add a custom agent, or edit ~/.jetbrains/acp.json:

{
  "default_mcp_settings": {
    "use_idea_mcp": true,
    "use_custom_mcp": true
  },
  "agent_servers": {
    "Open Interpreter": {
      "command": "interpreter",
      "args": ["acp"],
      "env": {
        "INTERPRETER_HOME": "/Users/you/.openinterpreter"
      }
    }
  }
}

Enable the IntelliJ MCP settings if you want Open Interpreter to receive IDE context and tools through the client.

VS Code

VS Code does not ship native ACP support, but community ACP clients such as the ACP extension use the same agent_servers shape:

{
  "agent_servers": {
    "Open Interpreter": {
      "command": "interpreter",
      "args": ["acp"],
      "env": {
        "INTERPRETER_HOME": "/Users/you/.openinterpreter"
      }
    }
  }
}

Use the extension's agent picker to start a session after saving the config.

What the Client Gets

Open Interpreter exposes:

• Session creation, closing, loading, and listing.
• Sandbox modes: read-only, workspace-write, and danger-full-access.
• Model and reasoning controls where the configured provider supports them.
• Streaming assistant messages and reasoning summaries.
• Tool progress for shell commands, file changes, MCP calls, web search, image operations, subagents, and other Open Interpreter tools.
• Permission requests for command execution and file changes.

The ACP client owns the UI. Open Interpreter owns model selection, provider transport, instructions, tools, approvals, sandboxing, and session state.

Troubleshooting

If the client cannot find interpreter, use an absolute command path:

{
  "command": "/usr/local/bin/interpreter",
  "args": ["acp"]
}

If authentication fails, start interpreter once in a terminal with the same INTERPRETER_HOME, finish provider login or API-key setup, then restart the ACP client.

If the client hangs, make sure it is launching interpreter acp directly over stdio. ACP clients should not launch interpreter without the acp subcommand, because that starts the terminal UI instead of the protocol server.