What Is T3 Code? 51,402 GitHub Stars on an Open-Source Control Plane (2026)

At 03:46:26 UTC on November 8, 2026, the GitHub API reported 51,402 stars for pingdotgg/t3code. The same response reported 9,893 forks, an MIT license, and an empty description field. The latest release was v0.1.7, published November 4. T3 Code is the project’s open-source control plane for coding agents you already run. The README’s own phrase is narrower: an “agent harness control surface.”

This page was published November 8, 2026. The homepage, the README, the FAQ, the install guide, the permission and remote docs, the ACP provider page, the license file, and both GitHub API calls were read that morning. Where those pages name different agents, the table below keeps them apart.

What is T3 Code?

T3 Code is a desktop, web, and mobile control surface for coding agents already installed on a machine. The homepage, read November 8, says “the open-source control plane for coding agents” and “bring your own subscription.” The README says it “enables control of the agents on your machine” from an iOS app, an Android app, a web app, and an Electron desktop app. The repository homepage field points at t3.codes. The MIT license file on main is copyright 2026 T3 Tools Inc.

It is not a new coding model. The FAQ’s cost answer says you bring your own subscription and still pay your provider. The homepage says T3 Code does not resell tokens. You plug in credentials you already have, and the app orchestrates those agents.

The repository was created February 8, 2026. The API’s pushed_at on the snapshot below was November 8, 2026 at 03:41:34 UTC, a few minutes before the star count was read. The default branch is main. The primary language field was TypeScript.

Is T3 Code a harness?

The FAQ answers that question with no. Read November 8, the “Is T3 Code a harness?” section says a harness is what turns an LLM into an agent: it gives the model tools, a way to call them, runs them, and hands the results back. It names Claude Code as an interface plus a harness underneath, the Anthropic Agent SDK, and Codex as the app plus the Codex App Server. The next sentence is the definition that matters here: T3 Code uses those harnesses. It is not a harness, and it does not work without one.

The same answer uses a kitchen line. The model is the chef. The harness is the kitchen. T3 Code is the restaurant that lets you bring whatever kitchen you want. That matches the definition on this site. A harness is the runtime around a model: the loop, the tools, the memory, the permissions, and the stop conditions. What is an AI agent harness? is the longer guide. T3 Code sits outside that runtime and drives it.

The homepage does not use the FAQ’s wording. Its testimonials, on the same page the same morning, include the line “Since I've started using T3 Code I no longer care which harness is used,” and another that calls T3 Code “the one AI harness.” Those are quotes the homepage prints. They are not the FAQ’s definition. The FAQ is the project’s own answer to the question, and the answer is no.

Which coding agents can it control, according to which page?

A single merged list would be wrong. On November 8, 2026, the homepage, the README, the FAQ, the install guide, and the ACP page named different sets. Kiro, Pi, Muse Code, and fx each appear on some of those pages and not on others. The release notes for v0.1.7 add one more line that the README’s install warning does not: “Kiro (beta). Use Kiro as an Early Access provider.”

The install table is the longest list in the docs read that morning. It also says how each one is signed in: codex login, claude auth login, Cursor’s agent login, grok login, kiro-cli login, opencode auth login, Google sign-in for Antigravity from T3 Code’s settings, pi then /login, muse login, and fx login. Antigravity and a ChatGPT-connected Codex can use a managed runtime without a binary on PATH. The others need the CLI on the server machine, which is the computer running T3 Code, not the phone.

The ACP page is a second door. It says T3 Code can run installed ACP commands, or coding agents from the official ACP Registry. Registry agents bring their own models, tools, and sign-in. T3 Code provides projects, threads, checkpoints, and task delegation. The page’s example of a local command is DeepSeek Harness: executable dsh, arguments --profile and acp. Registry agents are third-party code. The page says to review an agent’s source and license before adding it.

Claude Code, Codex, and Cursor are the three of those names with directory pages here. The November 5 framework snapshot is a different object: 19 agent-framework repositories, with their own star counts. T3 Code is a control surface for agents you already run. It is not one of those 19 rows, and this page does not reprint that table.

How much does T3 Code cost?

The FAQ’s answer is nothing. The heading “How much does T3 Code cost?” says there is no way to pay the project to use T3 Code. You bring your own subscription. It is free and open source (MIT). You still pay your provider the way you already do. The README’s version of the same point is the heading “Wait, what are you selling me?” and the answer “Nothing.”

T3 Connect, the built-in remote link, is a separate sentence. The FAQ says it is free too, for now. Every account can connect up to 3 machines running T3 Code. Phones and browsers that connect to them do not count. Need more machines, and the FAQ says to link a paid T3 Chat account. Pro raises the cap to 10 machines. Premier raises it to 100. The free T3 Chat plan keeps the normal limit of 3. The FAQ does not print a dollar price for Pro or Premier, and this page does not add one.

The same answer includes the hedge worth quoting: “We don't want to charge for this, but if the bill gets unacceptable we may have to change course.” The alternatives it names are forking the project, self-hosting, or using Tailscale. The remote-access doc repeats the 3, 10, and 100 machine caps and does not repeat the words “for now.”

What you still pay is the agent subscription. Claude Code vs Codex pricing is the dated comparison of those plans. This page does not restate those prices. The FAQ also says T3 Code can take an API key or an OpenRouter endpoint on a separate provider instance, so a subscription is the default path, not the only one the docs describe.

How does remote and mobile control work?

The agents run where T3 Code runs: a desktop, a server, or a VPS. You connect to that machine from somewhere else. The FAQ says less than 5 percent of that author’s prompts run on a laptop, and that npx t3 up on the machine is the remote path they use. The phone does not become the computer. The FAQ’s mobile answer says the iOS or Android app connects to your computer with T3 Connect or a pairing link, the phone is a remote, and the computer does the work and needs to stay on.

The remote-access page, read the same morning, names three ways onto the host. T3 Connect is the account link: sign in on the host, enable it, and pick that environment on the other device. t3 up asks whether other devices should use T3 Connect, the local network, or both. On Linux and macOS it then runs T3 Code as a background service. On Windows the background service is not available, so t3 up prints a t3 serve command to leave running in a terminal. LAN pairing uses a link or QR code from t3 pair or t3 up --lan. The page also documents Tailscale and other VPN routes, and a desktop mode that turns the local environment off so that computer only drives work running elsewhere.

The web client is app.t3.codes. The README links the iOS app as “T3 Code: Remote Claude & More” and the Android package com.t3tools.t3code. This page did not open the store listings for ratings or prices.

That split is the opposite of an agent that operates the screen in front of you. Can an AI agent use your computer? is the computer-use guide. T3 Code’s phone is a remote control for a machine that stays on.

One privacy sentence from the FAQ belongs next to the remote claim. Your code and prompts go from your machine to your provider, not to T3. With T3 Connect, traffic rides a Cloudflare tunnel on your machine. The FAQ says the relay sets up the link and does not proxy your HTTP or WebSocket sessions. T3 Code does send product usage events, such as provider, model, and token totals. Prompts and file contents are not in that list. T3CODE_TELEMETRY_ENABLED=false turns it off. That is the FAQ’s account, read November 8.

What permission and approval modes exist?

The permission-modes page lists four modes. You pick one in the message composer, and it applies to that thread. The initial default is Full access. Projects can override the environment default. An administrator can cap the modes on a machine with a managed policy.

Those modes are an approval gate on the thread, and a human in the loop when the mode still asks. The page is explicit that providers do not all implement the four the same way. Auto uses automatic review on Codex, Claude, Cursor, Grok, and fx. OpenCode and Antigravity, which the page says lack an equivalent, fall back to asking. Claude can still ask in Full access when its own settings disable bypass permissions. Antigravity can still send native approval requests in Full access.

Approve or reject in the conversation. The page says subagent approval requests appear in the parent conversation with the subagent’s name, so you can answer without switching threads. The v0.1.7 notes call that “child approvals.”

How is T3 Code different from Claude Code or the Codex app?

Claude Code and the Codex app are harnesses T3 Code can drive. T3 Code is the surface in front of both, plus the other agents in the table above. The FAQ’s “why use T3 Code” answer says the project was inspired by the Codex desktop app, Conductor, Claude Desktop, and Cursor Glass, and that none met its bar. What it lists instead: bring your own subscription, remote with npx t3 up, a mobile app, and source you can fork. The README adds that every agent thread writes to its own work, and the homepage says one button can open a GitHub pull request with a generated title, body, and changelog.

The practical difference is whose software is the loop. In Claude Code, you are in Claude Code. In the Codex app, you are in Codex. In T3 Code, the FAQ says those products keep doing the tool calls, and T3 Code is the place you start the thread, pick the permission mode, and attach a phone. Switching agents does not switch you to a new model sold by T3. It switches which installed harness the thread uses.

Two limits from the FAQ are part of that difference. First, T3 Code drives Claude through Anthropic’s Agent SDK. The FAQ says that path drains Claude’s 5-hour limit faster than using Claude Code directly, that weekly limits are unaffected, and that the project has reported it to Anthropic. That is the FAQ’s account of a bug, not a measurement this page re-ran. Second, the FAQ says T3 Code does not support local models. The sentence is blunt: locally runnable models are not capable of meaningful engineering work. A reader who wanted an offline model is outside the product the FAQ describes.

The cost difference is the license. Claude Code and Codex are paid through their own plans, compared in Claude Code vs Codex pricing. T3 Code’s FAQ says the control surface itself has no fee. The subscription you bring is still a subscription.

What do today's GitHub numbers say?

This table is the original snapshot. It is the response from GET https://api.github.com/repos/pingdotgg/t3code and GET https://api.github.com/repos/pingdotgg/t3code/releases/latest at 2026-11-08T03:46:26Z. GitHub’s open_issues_count counts open issues and open pull requests together. The homepage the same morning printed “51k+” next to the download button, and a separate line, “Tolerated by over 500,000 devs.” The star count and that dev line are different claims. Only the star count is in the API response.

The numbers in the table are licensed CC BY 4.0. Credit AI Agents Info if you reuse them. They are a reading of a public API at one minute, not a forecast. The November framework snapshot is the same kind of dated GitHub reading, for a different set of repositories.

v0.1.7 is not a draft and not a prerelease. The release name is “T3 Code v0.1.7.” The notes open with 659 changes from 160 contributors, then Kiro as an Early Access provider, managed policies, one-time schedules, child approvals, shareable thread links, usage filters, and model-switch history. The repository’s created_at is 2026-02-08T23:46:22Z. The description field in the API was null.

Stars on a fast-moving repository move after a snapshot. A count you see on GitHub later on November 8 can differ from 51,402. The row above is the count at the timestamp in the last row.

This guide is filed under harnesses, tools and SaaS, and how agents work.

Frequently asked questions

What is T3 Code?

It is an open-source control surface for coding agents you already have installed. The homepage’s phrase is “open-source control plane.” The README’s phrase is “agent harness control surface.” The GitHub organization is pingdotgg. The MIT license file is copyright T3 Tools Inc. The app does not sell a model of its own.

Is T3 Code a harness?

The FAQ says no. A harness gives the model tools and runs them. T3 Code uses a harness such as Claude Code or Codex, and it does not work if you have not installed one. The glossary entry is harness. The guide is What is an AI agent harness?.

Which coding agents can it control?

Read the page, not a merged list. On November 8, 2026, the homepage named six, the README named seven including fx, the FAQ named eight including Pi and Muse Code and not fx, and the install guide named ten including Kiro. The ACP Registry is an extra way to add agents the first-class list does not include.

How much does T3 Code cost?

Nothing, according to the FAQ. It is MIT-licensed, and you bring your own subscription. T3 Connect is free for now, for up to 3 machines. T3 Chat Pro raises that to 10, and Premier to 100. Phones and browsers do not count toward the cap. The FAQ names those plans and does not print a dollar price.

How do you use T3 Code from a phone?

Install the iOS or Android app and connect it to a computer that is running T3 Code. The phone is a remote. The computer stays on and does the work. T3 Connect, LAN pairing, and Tailscale are the documented routes.

What permission modes does T3 Code have?

Supervised, Auto-accept edits, Auto, and Full access. The documented initial default is Full access. Muse Code, Kiro, fx, and Grok do not offer every mode. The details are in the table above, from the permission-modes page read November 8, 2026.