I built a workbench for the AI coding CLIs I already use

le lohas — Fri, 28 Aug 2026 01:47:39 GMT

AI coding CLIs are already good at writing and changing code. The part that kept breaking for me was everything around the run.

A task might start in Claude, continue in Codex, move to another computer for a test, and end in a terminal window I could not find a week later. Adding another model did not fix that. I needed a way to keep the work reviewable and easy to resume.

That became Agent OS, an open-source desktop workbench for the AI CLIs already installed and authenticated on your machine.

It does not replace the CLI

Agent OS currently works with Claude, Codex, Gemini, Cursor Agent, OpenCode, Pi, Hermes, and OpenClaw.

You still use the original CLI and provider account. Agent OS does not bypass subscriptions, OAuth, API keys, or tool permissions. It adds a workbench around those tools:

structured sessions beside native CLI terminals
tasks with separate run records and deliverable review
side-by-side comparison across Web, CLI, and agent panels
one-off schedules and Cron
local or explicitly authorized remote runtime hosts
scoped memory and separate Markdown knowledge

A task should be reviewed, not merely finished

An agent can produce a confident final response while the actual work is incomplete. That is why a run in Agent OS can move into Review instead of being treated as done.

The user can inspect the available execution events, compare attempts, and decide whether the deliverable passes. When two agents take different approaches, their outputs can stay side by side instead of being copied between terminals.

This is a small distinction in the interface, but it changes the contract. The agent proposes completion. A person accepts it.

Memory and knowledge have different jobs

I originally treated durable context as one feature. That did not hold up.

Memory is compact, scoped context that may be recalled in a later agent turn. Knowledge is long-form Markdown written for people. Knowledge supports topics, tags, sources, drafts, publishing, local comments, and favorites. It enters a prompt only when someone explicitly selects it as a reference.

Keeping them separate prevents a useful article from becoming automatic prompt baggage.

Local-first, with explicit remote boundaries

Session metadata, tasks, indexes, preferences, memory, and knowledge stay on the local machine by default.

For remote work, another Agent OS desktop or a headless Runtime node can expose an approved set of agents, folders, and capabilities. The managed computer controls that scope and can revoke it. Pairing a runtime is not permission to use the whole machine.

The selected CLI and model provider still determine where model requests go. Agent OS does not pretend otherwise.

The current release

Agent OS v0.4.2 is available for macOS, Windows, and Linux under Apache-2.0.

The macOS build is not notarized by Apple. Download it only from the official GitHub Release, verify the published SHA-256 digest and source provenance, and follow the documented first-launch steps if macOS blocks it.

Links: