T3 Code

Detailed report

T3 Code supports agents running on your own machines, with control from laptops and phones. It combines remote execution, Git worktrees and review tools in one interface. You remain responsible for the hosts, provider integrations and credentials.

Product and evidence

The product is T3 Code at t3.codes, maintained in pingdotgg/t3code. The official repository links the site and identifies its logo asset as t3-black-web-icon-192x192.png. This is distinct from T3 Stack and the separate T3 Chat subscription.

Research date: November 13, 2026. Many official documentation pages have no publication date. Those are current repository documents retrieved on November 13, 2026. Features on main may precede a stable release.

What it offers

T3 Code lets you keep Claude Code, Codex and other agents under their existing accounts, then manage them through one desktop, web or mobile interface. You can change providers within a thread, delegate work between agents, and review changes before committing. It uses the providers’ underlying agent runtimes. It does not replace them with its own model or sell bundled inference. Official README

For a multi-device setup, the benefit is continuity: an always-on server owns the work, while whichever device is convenient becomes the controller. That could reduce the manual work of managing SSH connections, persistent terminals and copied context across devices.

Agents, models and payment

The current provider list includes Claude Code, Codex, Cursor, Grok Build, OpenCode, Antigravity, Kiro, Pi, Muse Code, fx and ACP-compatible agents. ACP is a protocol through which compatible agents expose their functions. Availability and integration depth vary. This is not a guarantee that every native extension or command works identically. Provider setup

There is no fixed T3 model allowance. Models depend on the configured provider and account. Codex can use a ChatGPT plan or existing CLI login. Claude supports its normal login, separate account configurations, API credentials and documented OpenRouter routing. Thus, existing subscriptions and bring-your-own-key configurations are both supported, through the relevant integration. Codex guide, Claude guide

Your machines, remote access and persistence

Desktop clients cover macOS, Windows and Linux. Phone clients cover iOS and Android. Linux desktop packages include x86-64 and ARM64 builds. Headless operation is explicit, with a user service available for Linux, macOS and Windows. Intel Macs have a documented limitation: the standalone executable is unavailable, although the desktop app exists. Installation, Background service

Three documented connection paths matter:

• SSH: the desktop app starts or reuses a server on a Linux or Apple Silicon Mac host and forwards its port. Code, credentials and execution stay there.
• Direct/private networking: LAN pairing and Tailscale connect clients to your server.
• T3 Connect: managed sign-in and tunnel setup avoids configuring router forwarding.

These are access routes to your compute, rather than vendor-provided coding sandboxes. Remote access

The environment owns durable state, files, terminals and provider processes. Its identity survives server restarts and address changes. Reconnecting clients return to that environment. This supports switching devices, but does not promise uninterrupted execution after a host crash. Existing threads remain on their originating machine rather than migrating automatically. Remote architecture

T3 Connect uses Clerk identity and managed tunnels. Its relay brokers access. Normal application traffic uses the environment’s tunnel hostname. The technical documentation explicitly treats the relay as a trusted component. Connect architecture

Isolation, orchestration and shipping

Threads can receive separate Git worktrees: separate working directories and branches sharing one repository. You can also submit one prompt to several models, each receiving its own thread and worktree. That separates code changes, but does not provide container-level security or automatically separate databases and ports. Thread workflow

Built-in source control supports diff review, commits, pushes and pull requests across several hosting services. Tests remain commands executed by agents or project scripts. The interface is not proof that changes passed meaningful checks. Source control

Outside agents can control threads through MCP, a standard tool interface. Linked environments also let agents launch, message, wait for and stop work on another machine. That capability is documented as beta. Outside agents, Peer links

Setup, security and teams

Setup means installing T3 Code, configuring providers on each execution host, and choosing a connection method. The server’s shell must be able to find the provider executables. Check this if remote setup fails.

The software is MIT licensed and can run on your infrastructure. That does not mean all data stays local: model requests still reach providers, and hosted connectivity introduces cloud dependencies.

Telemetry is enabled unless disabled. Product events include provider, model, token totals and hashed identifiers, but exclude prompts, responses and file contents. Server opt-out does not disable every browser/mobile diagnostic stream or separately enabled activity publishing. This distinction matters for company repositories. Telemetry documentation

Teams gain administrator policies for allowed providers and permission limits. Beta guest links allow a teammate to follow a thread and propose messages, subject to owner approval. Guests cannot inspect diffs. These are useful collaboration features, but the evidence reviewed does not establish a complete enterprise administration, compliance or support offering. Managed policies, Guest links

Pricing

T3 Code costs $0, with model subscriptions, API usage and your compute paid separately. T3 Connect is currently free for three linked execution machines. Phones and browsers do not count. Linking T3 Chat Pro raises this to 10, Premier to 100. Direct networking remains an alternative. The vendor explicitly reserves the possibility of changing Connect pricing. Official FAQ

T3 Chat prices are reported as US$8/month Pro and US$50/month Premier, with applicable taxes. Premier’s price is confirmed in a team reply. I could not independently verify the signed-in checkout on the research date. These plans increase Connect capacity. They do not supply T3 Code’s coding-agent subscriptions. October 28 pricing research, Official billing discussion

Development and user reports

Development is active: stable v0.1.9 shipped November 11, 2026, listing 198 changes from 31 contributors. A v0.1.10 nightly followed November 12. Release activity does not establish reliability. Stable release, Release history

Independent evidence is limited but relevant:

• Daniel Dewhurst, November 2: describes actual use on a headless Ubuntu VM from a MacBook and iPhone. He values remote attachments and the thread-inbox workflow, while discussing misleading completion states and operational costs. First-person account
• Issue #4597, July 27: reported a headless crash after roughly 23.5 hours. Maintainers closed it as addressed on August 20.
• Issue #25172, October: reported substantial memory growth. Follow-up measurements supported a fix, and it closed October 28. These are credible historical failures, not evidence that current stable retains them. Earlier crash, Later leak and resolution

Older criticism also needs checking: the current FAQ says excess Claude five-hour quota consumption was fixed in v0.1.7, and current docs include usage tracking. Neither broad enthusiasm nor broad dissatisfaction can be inferred from this small sample. FAQ, Usage documentation

Recommendation and trial

Choose T3 Code if you want control over your hosts, multiple providers and device flexibility. Avoid making it your sole unattended production workflow until you have measured recovery and resource use.

Trial it on one Linux server, two laptops and a phone. Run two isolated tasks, disconnect both laptops, approve from mobile, then deliberately restart the server and revoke one device. Check recovered history, interrupted turns, Git state, memory and provider billing over 48 hours. This tests continuity and recovery under realistic interruptions.