claudectl

claudectl is a local dashboard and supervisor for developers running several Claude Code agents on macOS or Linux. It brings session activity, approvals and spending into one terminal view, with an optional local LLM that makes decisions based on your preferences. It's open source under the MIT license.

The dashboard shows which agents need input, are working or have stalled, alongside context usage, resource consumption and subagent activity. You can approve requests, send input or terminate sessions from the same view. Spending controls enforce session or daily limits, while health checks flag retry loops, cost spikes and overloaded context. Completed sessions have post-mortem reports showing where work stalled and costs accumulated.

The local supervisor can approve or deny tool calls, pass summarized context between agents and start sessions for parallel work. It learns from your corrections and manual decisions, with separate preferences for different projects. Decision reviews show the reasoning, confidence and past examples behind an action. It works with Ollama, llama.cpp, vLLM and LM Studio through OpenAI-compatible endpoints.

Dependency ordering and coordination help agents share context and avoid duplicate work. A headless mode keeps supervision active without the dashboard, and peer-to-peer connections let instances delegate tasks and share learned preferences across machines. claudectl itself needs no API key or account and keeps monitoring and learning data locally; configured webhooks can send data out.