Running multiple Claude Code sessions in parallel speeds up large projects. Knowing what each session is actually doing is another problem. Conductor, a new open-source project from developer rmindgh, addresses this by letting one Claude Code session monitor and control others.
The system connects to Claude Code's remote control API (available in v2.1.92 and newer) via WebSocket, while also reading conversation logs in JSONL format (a structured log file format where each line is a self-contained record). From a single orchestrator session, you get real-time visibility into all active sessions, can send tasks between them, and chain work so one session's output feeds directly into another's next task.
Auto-Approval and Guardrails
The most practical feature is automatic tool-call handling. Claude Code pauses regularly to ask for manual approval before running certain commands. Conductor can auto-approve safe calls and block dangerous ones - force pushes, destructive deletions - without requiring human input each time. This works through Claude Code's PreToolUse hooks, which intercept commands before execution.
Under the hood, Conductor operates across four layers: Awareness (reading session logs), Decisions (risk-based approval logic), Guardrails (hook-based blocking), and Remote Control (WebSocket API). Thirteen MCP tools handle the integration across these categories.
Full control features require a Claude Max or Team subscription. Standard accounts get read-only monitoring only.
The Dependency Risk
The project documentation is direct about the main risk: "the Remote Control WebSocket protocol and hook blocking mechanism are based on Claude Code's internal architecture. They may change between versions." Conductor is built on an undocumented, internal API. Any Claude Code update could break it without warning.
Requirements are Python 3.10+, Node.js, the websockets and httpx Python packages, and Claude Code sessions started with the --rc flag to enable remote control. This is not a plug-and-play install - wiring the hooks correctly takes some setup work.
The project launched with three GitHub stars. It's early, it carries real fragility risk from that undocumented API dependency, and anyone adopting it should expect to do maintenance work when Claude Code updates. For developers already running parallel sessions who want some structure around it, there's nothing else quite like it right now.