Rome open-sources an agentic OS that stores agent work as git-tracked actions and apps

Rome AI Lab has opened the source for Rome, describing it as "the agentic OS for humans and agents". The repository at github.com/rome-os/rome was created on 23 August and is at 481 stars and 34 forks as of 7 September, all under the MIT licence.

The pitch, in the project's own words: "Most progress in AI comes from scaling models. Rome scales the other axis, the environment: the tools, workflows, memory, and interfaces an agent works within." A conversation with an agent produces work once. A Rome environment keeps the work as a named operation the next agent, and every human on the same instance, can call again.

Actions, apps, and the unit that compounds

The core abstraction is the action: a typed operation that agents, scheduled routines and the app's own code can all invoke. Actions are ordinary source files in a pnpm monorepo, so they sit in a git history, review as pull requests, and survive a model swap. Around them Rome adds agents (app-owned collaborators with their own tools), skills (plain-language procedures loaded when an agent needs them), hooks into the message and event lifecycle, and app-private database and file storage.

A Rome App bundles those pieces behind an app.yaml manifest and a purpose-built web interface. The project draws the line as "a workflow is a verb; an app is a noun": use a workflow for a one-shot task, build an app when the work needs a home with editable data, multiple actions and a persistent agent. Two public SDKs, @rome-os/app-runtime for the backend and @rome-os/app-web-sdk for embedded views, let anyone author one and publish it to the Rome App Store.

Runs on your machine, or on Rome Cloud

A single Docker script pulls the published image and starts Rome, binding its dashboard to http://localhost:7663 on the loopback interface only, with named volumes for state so re-running the script upgrades the container without wiping data. Telemetry stays off until an OTEL_EXPORTER_OTLP_ENDPOINT is set. Rome Cloud, currently in a preview open to guardians, provisions the same environment as a managed service.

The comparison table in the README places Rome against Grok Bot from xAI, Nous Research's Hermes Agent, Manus, and Wabi. Each of the four "compounds" something: Grok Bot the hosted bot state, Hermes the agent's own notes, Manus a machine and its files, Wabi a network of shareable apps. Rome, by its own account, compounds the environment itself: "executable, composable capability owned by you."

The gap it targets is real for anyone using coding or research agents in a client engagement. A chat session is a good place to ask for something once, and a poor place to run a review queue, a price tracker, or a morning brief that has to be there tomorrow. What agents produce today usually lives in a transcript that has to be re-read every time. Rome files that work as source code the next agent can reach for, which is what the "agentic OS" framing carries. Whether the framing survives a first real production workload is the open question. 481 stars in 15 days is enough to justify a real evaluation, and well short of the level where a team would build critical work on it.

Source

Primary source: rome-os/rome on GitHub. Star and fork counts from the GitHub REST API, retrieved 2026-09-07. Rome Cloud and App Store URLs from the project's website at romeos.cc.