heedle: self-hosted AI agent that shows you the program before it runs, seeking testers

vhalaharvi · reddit · 2026-11-07

A developer rebuilt his earlier AgentScript project as heedle, an MIT-licensed self-hosted personal AI agent whose core idea is human approval before anything executes.
You ask in plain English, Claude writes a small program (browse sites, fill forms, run a skill), and you see exactly what it will touch before it runs.
Safety design: each browser runs in a throwaway Podman container restricted to the sites named in the program; personal data stays in the OS keychain with Claude only seeing placeholders; sign-ins/codes are handed back to the user; previously approved forms replay without calling Claude.
Execution is orchestrated by Temporal locally, with every run committed to Git for auditability.
Status: early and rough. CI runs end-to-end on Linux/macOS with a stand-in model; bring your own Anthropic API key. Install on Apple Silicon via brew install --cask heedle-ai/tap/heedle && heedle up; Linux builds from source. The author especially wants a security review of the browser agent. Most code was written by Claude while he designed and reviewed every component.

More from coding & agent