Claude Code
How AgentDock detects, launches, and integrates with the Claude Code CLI.
Claude Code is one of the three coding-agent CLIs AgentDock can drive. AgentDock does not install, bundle, or authenticate Claude Code for you — you set it up independently, and AgentDock uses whatever is already working on your system.
Prerequisites
- Claude Code's
claudeCLI installed and authenticated, per Anthropic's own instructions. - No AgentDock-specific setup beyond that — see Getting started: connecting your agents.
How AgentDock detects it
AgentDock searches your PATH (plus a few known install locations on Windows and macOS) for a runnable claude executable. Linux detection is PATH-only, with no known-install-directory fallback. If detection fails, see Troubleshooting: agent CLI not detected.
You can override the detected path in Settings → Agents, which also runs a live "Test" action and shows diagnostics (resolved path, executable type, version).
How the integration works
AgentDock talks to Claude Code through @anthropic-ai/claude-agent-sdk, keeping one persistent SDK session open per AgentDock session rather than re-spawning the CLI on every message. This is what makes live model and permission-mode switching possible mid-conversation.
Permission requests go through the SDK's real canUseTool callback (over --permission-prompt-tool stdio). When Claude Code wants to run a tool that isn't already allowed, AgentDock shows an inline interaction card in the conversation — not a separate modal — and blocks until you respond. See Permissions for the full list of Claude's permission modes.
Note
Model and permission-mode switching happen live, in the same session — you don't need to start a new session to change either.