T3 Code: Run Coding Agents From One Workspace

Run coding agents in one workspace with T3 Code. Set up providers, isolate tasks with worktrees, connect your phone and understand nightly customization.

Published

T3 Code: Run Coding Agents From One Workspace

T3 Code lets you run your existing coding agents from one workspace, keep separate tasks on separate branches, and pick up the same work from your phone. Add it when managing agents, changes and devices has become a job of its own.

The useful change is organizational: you can move from task to diff to pull request without rebuilding that workflow around each agent's interface. T3 Code is free and open source. Your agents still bill through their own plans.

Version scope, November 16, 2026: the stable baseline is v0.1.9, released November 11. This is a documentation-based guide, not a hands-on installation or phone-pairing test. Setup details follow the repository's current documentation; the newer agent-driven interface customization is explicitly marked as nightly below.

What T3 Code Adds to Your Existing Agent

T3 Code gives several agents a shared place to work. It is an open-source project from Theo Browne's team, with desktop apps for macOS, Windows and Linux, mobile apps for iOS and Android, and a web client.

Think of it as a workshop with several specialists at separate benches. The specialists are the coding agents. T3 Code organizes their workspaces, conversations and review process. It uses the agents' existing harnesses, the software that gives a model tools and runs its actions, rather than supplying a new model of its own.

The repository's feature overview describes model changes within a thread, delegation between agents, worktrees, diff review and remote access. A thread is a conversation attached to a task. A diff shows exactly which lines changed.

The reason to add it is that combination across providers. Worktrees or remote access alone are not proof that every agent's own app lacks them. If your current setup already handles one agent and one task comfortably, another interface needs to earn its place.

For choosing the underlying agent, use our AI coding agent comparison. For choosing where to edit code yourself, see our editor comparison.

Install It and Run a Small First Task

Start on the computer where the code and agents will live. Installing the phone app comes later.

  1. Install the desktop app. Use the official downloads. On macOS, the package-manager command is brew install --cask t3-code; on Windows, winget install T3Tools.T3Code. Linux has AppImage and .deb downloads. Choose the file for your machine's architecture. An AppImage must be made executable before launching.
  2. Connect one provider. Open Settings → Providers, select the environment, meaning the machine doing the work, and enable an agent using the table below. Its installation and login belong on that machine.
  3. Add an existing Git project. Git is the version-control system that records changes. In the web app, use Add project; on desktop, you can drop a folder onto the sidebar. With the desktop app running, t3 app also opens the terminal's current directory as a project.
  4. Start a thread and check its destination. Choose the project, model and permission mode. For an isolated code change, choose New worktree and its base branch before sending the task.
  5. Give it something easy to verify. For example: “Add a test for this existing validation rule, run the relevant test command, and explain the diff.” Inspect the changed files and test output before publishing anything.

For a web-first trial, npx t3@latest needs Node.js and runs the stable version selected by npm. It is not the way to invoke an installed nightly build. The installation guide also covers Linux launch details and Windows Subsystem for Linux. With WSL, install the provider tools inside the selected Linux distribution.

Connect Each Supported Agent

The common rule is simple: sign in where the agent runs, then enable it in T3 Code. A login on your laptop does not configure an agent on a remote server.

AgentSetup on the host machine
Claude CodeInstall it, then run claude auth login.
CodexUse Connect with ChatGPT in provider settings, or install the CLI and run codex login.
CursorInstall Cursor CLI and run agent login. Its executable is named cursor-agent.
OpenCodeInstall it and run opencode auth login.
Grok BuildInstall its CLI and run grok login.
AntigravityInstall and sign in with Google from T3 Code's provider settings.
KiroInstall Kiro CLI and run kiro-cli login.
PiInstall Pi with Node.js 22.19 or newer, then run pi and use /login.
Muse CodeInstall it on the server, run muse login, then enable the provider.
fxInstall it using its installer, then run fx login.

CLI means command-line interface: a program you can run from a terminal. If an installed agent is missing, set its Binary path in provider settings to tell T3 Code where that program lives. Check the provider card for compatibility warnings and recommended versions.

For another agent using ACP, the Agent Client Protocol that connects agent software to clients, open Settings → Providers → Add provider → Show more. Select a Registry agent, add it and complete its sign-in. Local ACP command instead connects an executable already installed on the host. Available models and options depend on what that agent exposes. See the ACP setup guide.

Use Worktrees to Separate Tasks, Then Review the Whole Branch

A worktree is a separate working directory attached to the same Git repository. It lets one task change its files while another task works elsewhere, without both editing the same checkout.

For a concrete example, give a checkout bug and a documentation update separate threads. Select New worktree for each, starting from the intended base branch. Each task gets its own branch and directory. You can then review each result independently.

Architectural flow from a main project to separate Task A and Task B workspaces, then review and merge.
Separate workspaces keep tasks apart. Review remains the gate before merging their changes.

The thread workflow also supports sending the same prompt to several models: Shift-click models in a new thread's picker, choose a base branch and send. Each selection starts a separate thread and worktree. This requires a Git project.

Use the turn's diff to inspect that turn's changes, then the branch comparison to inspect everything against the base branch. Run the project's relevant checks. When the change is ready, the thread's Git actions can commit, push and create a pull request, the proposal that someone reviews before merging code.

Configure Git and your code-host authentication on the machine running T3 Code. After signing in, use Settings → Source control → Rescan. The source-control guide explains authentication and review controls.

Worktrees separate files; they do not make conflicting design decisions compatible. Two tasks can still change the same behavior differently. Nor should you assume separate directories isolate databases, credentials or external services. Those need their own project setup.

The business consequence is fewer places to reconstruct a task's state: its conversation, files and review sit in one workflow. That is a concrete change in steps. There is no verified hours-saved figure here, and running more agents does not remove the review workload.

Continue From Your Phone While the Host Keeps Working

Your phone controls the host computer. The host still holds the project, credentials and running agents.

  1. On the desktop host, open Settings → Sharing & access, sign in and enable T3 Connect for that environment.
  2. Install the mobile app from the iOS or Android link on the official site.
  3. Sign in to the same T3 Connect account on your phone and select the environment.
  4. Open the existing thread to continue the conversation or handle an approval. Before relying on it away from your desk, check that you can reach the host from the network you intend to use.

For a command-line host, t3 up offers T3 Connect, local-network access or both, and runs T3 Code as a background service. If you rely on the desktop app's local server instead, closing its last window stops that server. The computer must remain running and reachable.

Desktop and phone connect to a host computer containing the coding agent and project files.
The phone is another control surface. The host must stay running and reachable.

The normal T3 Connect account limit is three host machines; phones and browsers do not count toward it. Direct pairing over a local network or Tailscale is another option. A pairing URL is a credential, and default pairing grants broad permissions. The remote-access guide covers narrower access and revocation.

For jobs that continue over a longer period, our guide to long-running production tasks addresses the separate question of supervising the work itself.

Interface Customization: Stable Themes, Nightly CSS

The new customization goes beyond choosing colors: an agent can restyle the interface using CSS, the rules that control how a web interface looks.

Keep the release boundary straight. Stable v0.1.9 has palette editing, theme import/export and environment themes in its versioned appearance guide. Agent-driven color and layout customization is listed in v0.1.10-nightly.20261114.4669, a prerelease from November 14, in the release history.

On a build with the new feature, open Settings → Appearance → Create theme, then Custom CSS. The CSS travels with the exported theme. Alternatively, ask an agent running on the server's machine to make a specific change, such as a taller chat header. That thread must use Full access.

There are useful boundaries. Mobile has its own appearance preferences. Environment themes do not apply to app.t3.codes and other connections. CSS class names can change between releases; the current guide recommends targeting stable-looking data- attributes instead. External file loading through theme CSS is restricted.

If a customization hides controls you need, choose another theme or run t3 theme set t3-chat on the server. The current appearance documentation explains the feature and recovery path. My advice: make one readability change first, then verify the diff and approval controls remain visible.

Five Workflows Worth Trying, Ranked

These are practical uses you could try, not measured customer results.

Priority and readerExact workflowWhy it could pay off
1. A small product team with unrelated fixesGive each fix its own thread and worktree; review and publish the branches separately.It keeps one unfinished task from occupying the checkout needed for another.
2. A builder already using several agentsSend the same bounded task to selected models in separate worktrees, then compare the diffs and checks.You can judge the implementations against the same acceptance criteria. Extra runs still use provider allowances.
3. A maintainer away from the deskStart a task on an available host, then answer its next question or approval from the phone.Work need not wait solely because the maintainer has left the keyboard. The host still needs connectivity.
4. An agency moving between client repositoriesKeep each project and task explicit, with its own review before creating a pull request.The workflow makes the destination of a change easier to inspect before it ships. It is not a substitute for account and data isolation.
5. A developer struggling with the interfaceOn the appropriate nightly, use a theme or a narrowly scoped CSS change to improve spacing and readability.A frequently used workspace can fit the person reviewing the code. Maintaining that customization is the catch.

Two Things Worth Building Around It

The strongest opportunity is a repository-specific worktree setup kit. A team could pay for a repeatable way to prepare dependencies, local configuration, test databases and development-server ports for each task directory. The smallest useful version would support one team's repository, check prerequisites and provide a cleanup command. T3 Code would remain the interface used to start and review the work.

Interest in the underlying workflow is not proof of demand for this exact product. The commercial test is whether a team repeatedly loses time making a fresh worktree usable. The catch is repository variation: a generic wrapper is easy to copy, while reliable setup knowledge is specific to each stack. That specificity is also why this is the strongest opportunity.

A second option is a provider onboarding and compatibility kit. An agency or internal tools team could use it to check which agent tools are installed on each host, guide authentication without collecting credentials, and document a known working configuration. An MVP could be a local diagnostic report plus a team setup guide, delivered as a small implementation service before becoming software.

Discovery interest is not proof that people want a paid setup kit. T3 Code already has provider settings, compatibility warnings and update controls, so duplicating those is a weak product. The opportunity would have to be the team's reproducible environment and onboarding process. Provider changes create an ongoing maintenance obligation.

When Adding T3 Code Is Worth It

Add it when you need a common workflow across agents, several isolated tasks, or access from another device. Stay with your existing setup when none of those solves a recurring problem.

The constraints are practical. Provider authentication and usage limits remain. A remote host must be available. Worktrees still need dependencies and tests. A generated pull request description is not a code review. Current-main documentation can describe features newer than the stable download, and nightly customization adds another moving part.

For your first workday with it, take one low-risk change through the complete loop: connect one agent, open the repository, create a worktree, run checks, inspect the diff and prepare the pull request. Add phone access only after that works. Keep T3 Code if the complete workflow is easier to manage.

Should I use T3 Code Nightly?

Use stable as the baseline for an initial setup. Nightly is a prerelease and is relevant if you specifically want the newer agent-driven CSS customization. This guide distinguishes stable v0.1.9 from the November 14 nightly that lists that feature.

Does T3 Code Mobile run agents on my phone?

The phone connects to a T3 Code server on another machine. That host runs the agents and holds the project. Keep it running and reachable, then connect through T3 Connect or a suitable pairing route.

Does T3 Code work on Windows and Linux?

Yes. Windows has a desktop installer and package-manager options. Linux has AppImage and Debian packages for supported architectures. If you use WSL, configure the agents inside the selected distribution.

Can I use T3 Code in a web browser?

Yes. The web client connects to your T3 Code server. The hosted app.t3.codes client needs an HTTPS endpoint; it does not make an unreachable host reachable by itself.

If your team wants a repeatable setup for agent workspaces, we can help build the production workflow around it.

Published
Category
Build
Related Articles
GitLab vs GitHub

GitLab vs GitHub

Choose GitLab or GitHub by hosting, CI/CD, coding agents, security and migration work, with current plans and a practical switching checklist.Nov 16, 2026Build
Chipp AI Review

Chipp AI Review

Chipp AI reviewed for consultants and agencies: branded agents, knowledge sources, payments, current plans and the limits that matter before you sell.Nov 14, 2026Build
Cursor Composer 2: What It Is and When to Use It

Cursor Composer 2: What It Is and When to Use It

What Cursor Composer 2 is, how Composer 2.5 improves it, and when to choose Composer, Claude or GPT for edits and architectural changes.Nov 14, 2026Build
Supabase Alternatives in 2026: Appwrite, Neon, Firebase and More

Supabase Alternatives in 2026: Appwrite, Neon, Firebase and More

Compare seven Supabase alternatives by database, backend services, self-hosting and migration work. See which fits, and when to stay on Supabase.Nov 14, 2026Build
Google Antigravity vs Claude Code

Google Antigravity vs Claude Code

Google Antigravity vs Claude Code: choose by models, parallel agents, permissions, CI and team controls, with current published plans and limits.Nov 13, 2026Build
Cloudflare vs Vercel in 2026

Cloudflare vs Vercel in 2026

Cloudflare vs Vercel for Next.js, static sites and AI backends: current runtime limits, a worked bill, spending controls and migration trade-offs.Nov 13, 2026Build
Resend Email

Resend Email

Resend email reviewed for product teams: API, SMTP, templates and delivery workflows, plus the production limits that change the choice.Nov 13, 2026Build
Best IDE for Claude Code in 2026: VS Code, JetBrains, Cursor, Zed and the Terminal

Best IDE for Claude Code in 2026: VS Code, JetBrains, Cursor, Zed and the Terminal

Choose a home for Claude Code by diff review, file context and parallel sessions. Compare VS Code, JetBrains, Cursor, Zed, tmux and Desktop.Nov 13, 2026Build
Newsletter

One letter, every Sunday.Working systems, not hot takes.

Weekly. No spam. Unsubscribe anytime.