T3 Code Reaches 38.1k Stars in Alpha, and Its Remote Access Docs Say to Treat Pairing Links as Passwords

T3 Code, an MIT-licensed app that drives coding agents already installed on your machine, now has 38.1k stars, 8.9k forks and 1.5k open issues on its GitHub repository. Its README still reads: "We are very very early in this project. Expect bugs."

The repository sits under the pingdotgg organisation. The repo page names no individual authors. A write-up on flaviocopes.com, updated November 1, names Theo and Julian as the developers and says the project passed 20,000 stars on August 27. That is 18,100 stars in about nine weeks, and stars measure interest, not use.

What T3 Code actually is

The README calls it an "agent harness control surface". It does not run a model. It starts the Claude Code, Codex, Cursor, Grok Build, OpenCode, Antigravity or fx tools you have already signed into, and shows them in one interface: an Electron desktop app, a web app, or an iOS and Android app.

The Flavio Copes write-up describes the architecture. A local server owns projects, threads, terminals and Git, and clients connect over one authenticated WebSocket. State is stored as events in SQLite. Each thread can work in its own Git worktree, and checkpoints use hidden Git refs.

The same write-up says the direct provider command-line tool is still the simplest choice for one task, and the quickest way to a new provider feature.

Releases since October 15

The release page shows how fast it moves. Version 0.1.0 shipped on October 15, per the write-up, and the page lists v0.1.6 as the latest stable release on October 31.

Build	Date	Notable change
v0.1.6	Oct 31	Side chats, 1Password for credentials, tokens kept out of traces
0.1.7 nightly 4016	Nov 1	One-time scheduled tasks, multi-PR linking
0.1.7 nightly 4035	Nov 1	Agents open a dev server in the preview
0.1.7 nightly 4048	Nov 1	Organisation-wide policies for approvals and updates

None of the visible release notes is labelled a security fix and none links a CVE. The v0.1.6 notes do include "Connection tickets and OAuth credentials stay out of traces", and the page lists related fixes that stop API keys and clone-URL tokens from reaching logs.

What the remote access guide warns about

The feature that draws a phone user is reaching an agent on a home machine from elsewhere. The remote access guide offers T3 Connect, LAN or Tailscale pairing, SSH hosts, and a hosted web app at app.t3.codes that connects straight to your server.

The guide says in so many words to treat pairing URLs and codes as passwords. It lets a host revoke pairing links, and a --scope flag narrows what a token can do, from filesystem access to terminal operations. Each account can link up to 3 machines through T3 Connect, and a paid T3 Chat plan raises that limit.

One line stands out. Running as root on a Linux host needs T3CODE_SERVER_BROWSER_SANDBOX=0, which turns off Chrome's sandbox for the remote browser tabs.

The Copes write-up adds that full access grants full access on the server machine, so permission modes do not make commands safe. It also says checkpoints are not backups, and that the usage page covers only Codex, Claude and Grok.

What readers should check before installing

The maintainers say they are "(mostly) not accepting contributions yet" and will consider small fixes, not large features. A bug found today therefore depends on the core team fixing it.

An agent harness that holds provider logins is a target. The tampered Claude Code extension in Cursor's marketplace left an SSH key behind in eleven hours, and the Claude-Mem repository shows how a fast-growing agent tool can carry licence questions its stars do not reveal.

1. August 2720K stars
2. November 238K stars

Source: flaviocopes.com (Aug 27 milestone) and GitHub repository page, accessed 2026-11-02

The next stable build after v0.1.6 is the 0.1.7 line, whose nightlies already carry the organisation-wide policy controls. No stable date has been published.