Omnirush released a desktop coding agent that uses frontier models on its own account, 1,866 stars in 15 days
Omnirush is a desktop coding agent for macOS and Linux that routes work through the vendor's own account on GPT 6 Astra by default, so a user signs in once and runs prompts without a model key or a per-request bill.
Image: Omnirush
Why it mattersA desktop agent that bills the vendor instead of the user removes the account step that stops casual developers from trying one; whether that stays free in a year is the question every reader should ask before taking a dependency on it.
A desktop coding agent that reaches top AI models without the user holding a provider key is unusual. Omnirush released one on 21 September and took it to 1,866 GitHub stars in the fifteen days since.
The project is omnirush-ai/omnirush-gui, published under MIT and shipping as a native app for macOS on Apple Silicon and Linux on x64 and arm64 with .deb, .rpm, .pacman, AppImage and tarball builds. Omnirush's README calls the product "a desktop coding agent with free access to frontier models." The site title phrases it the same way: "State-of-the-art models, free, for every profession." The latest release at the time of writing is v3.1.3, published on 5 October.
How the account step works
The app ships with no API keys and no model subscriptions on the user's side. A user clicks sign in, approves a device code at omnirush.ai/console, and the app is ready. Omnirush's README states that GPT 6 Astra is the default model and that GPT 6.1 Sol and the rest of the account's catalog are available from the model picker. The user can also plug in their own OpenAI, Anthropic or other provider key if they prefer, and that path uses the user's own account for billing.
The service tier that pays for the default route is Omnirush's own. The repository's docs/features/managed-models-dpa.md describes a managed-models path with an eligibility flag per organization and an audit entry when an admin toggles it. The free offer is Omnirush's call to make or revoke, so a reader should treat the access as a vendor decision that can change the day the subsidy runs out.
What the agent itself does
Beyond model routing, the agent runs git workflows and tool calls. Omnirush's README states that it clones, branches, creates worktrees, commits, pushes, and opens and reviews pull requests through the gh command line. Write commands prompt once per session; destructive commands always prompt. A full-permissions toggle in the composer flips between asking for every call and allowing them. Skills and MCP servers added from the in-app library are shared across every model in the picker, so the same tool set follows a chat when the user swaps the model.
The project is a monorepo with its own evals, scenarios and worlds directories, plus enterprise documentation for AWS EKS, Azure AKS and GCP GKE deployments. Updates ship through the app, with the macOS version downloading a new DMG and the Linux installs updating in place.
The honest limit on the free claim
Omnirush's own words frame the model access as free. Nothing in the repository defines "free" as a monthly limit, a token allowance, a trial length or a paid upgrade path, so a reader deciding to build on it has nothing to count against. The traction is real at 1,866 stars across two weeks, and the enterprise paths above suggest a paying tier sitting alongside the free one, but the free side has no stated floor. Pick it up this week to see if it fits a workflow; wait on a dependency that assumes it stays free.
Source
- Primary: Omnirush, omnirush-ai/omnirush-gui on GitHub, MIT, 1,866 stars on 6 October 2026, latest release v3.1.3
- Product site: omnirush.ai
This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.
Get AI News in your inbox
New developer tools, model and agent releases, and how teams are actually using them to release software. Short, and only when there is something worth reading.