Approving on your behalf, too

Running 5 to 8 Claude Code sessions at once made dismissing approval popups one by one a real burden, so I built claudectl as a solution.

This tool is a TUI dashboard that sits on top of multiple sessions, where a local LLM (ollama/llama.cpp) makes the approval decisions instead.

The core is learning from user corrections.

• Press b to approve an action the brain proposed, B to reject it
• All corrections are saved as logs
• With just about 10 decisions accumulated, patterns are distilled such as "always allow cargo test" or "forbid rm -rf outside /tmp"
• Tracks per-tool accuracy, and if it keeps getting things wrong especially on Bash, it automatically raises the confidence threshold further
• Once 50+ decisions have accumulated, it follows the user's tendencies fairly well, weighting rejection signals 8x stronger than approvals

There are also plenty of additional features.

• A dashboard showing all active sessions
• Monitoring states like loop detection, stalls, cost spikes, and context saturation
• Detecting file conflicts across sessions
• Multi-session orchestration that accounts for dependency order
• Recording highlights as GIF/asciicast
• Approving permission prompts right within the dashboard

The setup is local execution, no cloud API, no telemetry, and decision logs and preferences are stored in ~/.claudectl/brain/. It can be installed via brew install mercurialsolo/tap/claudectl, and the brain feature can be turned on with claudectl --brain; the author states it was implemented in Rust.