Hey Product Hunt 👋 I’ve been building Agent CLI as an open-source experiment in a question I kept running into while using AI agents: How much of an agent should we actually have to trust? A model can say it ran a test, changed a file, recovered from a failure, or finished a task. But for local agents with real machine access, I wanted those things to exist as system state rather than narration. That became Agent CLI: a local-first, terminal-native agent harness where actions go through one explicit policy boundary, execution becomes append-only evidence, workspace changes are reversible, verification comes from real processes, and a crashed session can reconcile and resume from what actually happened. Over time the same runtime grew to support five model providers, bounded multi-agent work, persistent memory, web research, browser verification, DOCX/PDF workflows, Git checkpoints and explicit GitHub delivery. Agent CLI is free, MIT-licensed and fully open source. There’s no paid product behind this launch — you clone it locally, bring your own model API key, and the state stays on your machine by default. I’ve also tried to document the limits rather than hide them. For example, the OS-enforced automatic-command sandbox is currently Windows-only, approved commands run with the user’s normal authority, and there is no network confidentiality boundary yet. The V1 development cycle is now essentially complete, so I’d rather put it in front of people outside my own test environment than keep polishing it alone. I’d especially love feedback on three things: 1. Does the evidence/verification model give you useful confidence, or just more information to inspect? 2. Are the approval and authority boundaries understandable during real use? 3. What is the first workflow or edge case you would try to break? Thanks for taking a look — technical criticism is very welcome.

Agent CLI - A local-first AI agent harness that can prove what it did

Replies