T3 Code guide
How T3 Code works, feature by feature. I put this together from its docs, its release notes and Theo’s videos.
Looking for a setting? Settings reference
Best ways to use T3 Code
Seven habits from Theo’s own videos. Open one to see what he said.
Spawn worktrees by default and fire prompts off in the backgroundA worktree is a separate copy of your project's files so agents can work without clashing.
Give every task its own branch and directory so parallel work stays manageable. Press Cmd+Enter instead of Enter to launch the prompt as a new thread while you stay on the composer, then brain-dump the next job right away.
By default, I highly recommend that you spawn work trees because it just makes managing everything way easier when you're working in parallel, but that's your choice.
And here is where one of my favorite T3 code pro tips comes in. I'm not going to press enter, I'm going to press command enter, which sends off the thread in the background and leaves me exactly where I am. So I can now send off another prompt without having to do anything. It's very nice.
So if I I don't know say I want you to work on feature one. I could hit enter and it would send and start the thread the normal way or I can hold down command and press enter and it opens a new thread and I stay here.
T3 Code docs: thread sidebar
Steer running agents instead of queuing follow-ups
Set follow-ups to steer, then send quick corrections or extra requests mid-run, such as telling the agent you already ran an update or asking for findings as HTML for your phone. The agent adjusts course without stopping or restarting.
First off, I noticed it being confused because it was checking the history and saw that the instance was updated. So I responded when it was running with a quick steer saying I manually ran the update after so that it knew the update had been run.
PRs that you care about or what I'll do here. Put your findings in HTML so that I can more easily look at them on my phone. This is now going to go in as additional steering context, and I…
…have steering too, so when the model's annoying you, you can send a message to do a follow-up. We also have queuing, so queued message, and you can have a hot key to swap between the two. I don't like queuing, so I set my default to be steer, but you can change that in settings.
4 more ways
End prompts with "file a PR and babysit it"
Ask the agent to implement the change, open a pull request and watch it, so it reacts to CI failures and review comments on its own. Have it attach a video proving the fix works, and tell it in your PR skills to use the built-in watch feature rather than its own polling.
What would it look like to cache that similar to how we cache on the sidebar? If it's relatively easy, I would like you to implement it and file a PR and babyset it.
This is a skill that gets invoked when I tell the model to watch or babysit a pull request. The point of this skill is to keep the model checking for updates, for new comments, and make sure by the time I go check the thread, it's good to go.
…is again cuz my babysit PR skill was leading to OpenAI models not taking advantage of T3 Code's built-in watchers and doing stupid instead. So, I added another section here. In T3 Code, you should use the watch pull request feature. If you stop before the merge, call unwatch pull request first.
Or just tell it to throw a video do the pull request so you know it worked. That's what we do most of the time now with T3 code. Here's a PR from Maria.
Run agents on a fleet of other machines and drive them remotely
Set up T3 Connect on spare boxes and servers so threads run there rather than on your laptop. Let auto load balancing spread new threads across machines, and kick off work from your phone by voice, checking results later at your desk.
Connect is set up on this computer, which means I can quit from SSH, clear, never think about it again, and manage the whole machine all from my desktop app or my phone app. It is genuinely the most pleasant way I've ever found to manage all of my AI coding stuff, and I basically never run threads on my own computer anymore.
…been around for a bit know, I have moved pretty much all my agent work off my laptop and onto other machines, mostly my apartment, but a few remote servers as well. So normally when I use T3 Code I trigger it somewhere else.
Funny enough, I sent this one from voice to text on my iPhone as I was getting my hair ready in the bathroom before stream because I wanted to have it ready to use if I wanted to demo it during stream. So I was on my phone.
And if you're too lazy to even pick which server to run, Maria has built an awesome feature for you in settings here. Load balancing. You can now set up auto load balancing across all of the servers that you have connected in T3 code, so it will fire off the threads across your different machines.
Hand the agent a screenshot of the bug and say "Fix it"
Skip working out the fix yourself: paste the screenshot and ask it to fix, test, record a proof video, link the PR and babysit review. If it fails, you know the problem needs more of your own thinking.
So, now I just hand the screenshot to the agent and say, ‘Fix it.’ And it often does. And if it fails to, I now know this problem is too hard for the agent to do itself. At which point, I know it's time to think more about it.
…get a DM on my phone from a bug a user had in T3 code. So, I screenshot it. I paste it to my agent in T3 code on my phone and say, ‘Fix this, test it, record a video showing it works now, and link me the PR when you're done. Babysit it until all the issues that come up in review are addressed.’
I'll still screenshot the tweet and go post it into T3 Code myself. I won't have the agents propose fixes based on what people are tweeting yet.
Have Claude call Codex through T3 Code to review its work
Connect agents to the T3 Code MCP so one can spin up threads with another. Theo has Claude ask Codex to review Claude's changes because he finds OpenAI models more thorough at analysis, and lets Codex start Claude threads to take tasks from email to shipped.
…because Open AI models I've still found are a bit more thorough with their analysis and will find things deeper than Anthropic models tend to. So, I do a wield and the vast majority of my Codex usage is either me using it to like do my email and computer you or it's coming from Claude calling Codex via T3 code in order to review Claude's work.
Now that I've connected the T3 code MCP to my Codex, I can have Codex and dots spin up Claude threads in T3 code. I've created the pipes to allow these things to connect and connected the data that my agents need to operate. And now I can go from it tracing an email for me to shipping real changes with other agents and other programs across my systems.
T3 Code docs: outside agents
Getting started
Setting up the machine where your agents will run, such as a desktop, server, or VPS, so other devices can control it.
Opening projects from a terminal with t3 app, or by dragging folders onto the sidebar.
Getting at least one ready provider so you can start a thread.
Approving steps and following threads from your phone after starting them at your desk.
Keeping a host available for remote access from phones, browsers, and other machines.
Opening a new thread for the current directory from a terminal with t3 app, which adds the project if needed.
Keeping app and server versions compatible. Mismatched protocol versions refuse to connect.
Letting your own coding agent investigate a T3 Code problem.
Threads
Keeping separate tasks apart. Choose New worktree when the code changes need their own branch and working directory.
Quick one-off jobs, such as converting files, that do not belong to a repository.
Keeping your most important threads at the top and arranging active work in the order you want.
Clearing finished work from your active list. Settling also closes the thread's idle terminals while keeping their output.
Jumping back to earlier work, and sharing thread links that open the thread or paste into a composer as a thread chip.
Asking clarifying questions about ongoing work without cluttering or interrupting the main thread.
Continuing interrupted work without losing your place or your queued messages.
Sharing a prompt on social media, or putting a one-click fix in a pull request comment.
Continue the same thread with another provider. For longer conversations, T3 Code favors recent requests and answers, the original request, and relevant activity such as command outcomes. Selected messages keep their full text, order, and user or assistant role.
Avoid manual compaction when a long thread moves to a model with less room. Changing models or context-window options discards stale usage and compaction thresholds.
Keep working on old conversations in V2 without running an import command. Stashed prompts, drafts, layout, and theme also come over on the desktop app's first launch.
Parking a thread you want to return to at a specific time.
Messages and context
Sharing big logs or files with the agent without filling the model context, while still letting it inspect the content.
Writing structured prompts with lists, task lists, quotes, headings, and code blocks.
Giving the agent files and images to work with.
Adding instructions mid-turn or lining up the next tasks without waiting.
Telling the agent exactly which words of its earlier answer you are responding to.
Retrying a task with a better prompt, optionally restoring workspace files to their earlier checkpoint.
Reusing earlier prompt text or parking a draft while you do something else.
Long tasks with a clear finish line, such as getting all tests in a package to pass.
Seeing data or designs at a glance. Pages use your current theme and follow light and dark mode. Scripts run in a sandbox away from T3 Code and your session, and links you click open in your browser.
Understanding slow first previews and failed external resources. Previews can load public resources but block names that resolve to private or reserved addresses, which fake-IP DNS proxies can return.
Working with tool results in an app that follows your theme. An app can call tools on its own server and post to the thread, and it can tell the agent about choices you made with your next message.
Giving the agent screenshots, logs, or documents as part of an answer. Files upload to the environment running the thread, and the agent receives paths it can open with its tools.
Dictating into your draft with the composer's microphone. Navigating to another screen keeps dictation active.
Controlling which local speech model turns your dictation into text.
Transcribing with a cloud provider. Other devices with dictation access to the environment can use the saved key, but only the device you set up opts in.
Reaching skills quickly by typing a slash, or decluttering the menu if you rarely use them.
Making dictation use the microphone you want, such as preferring the built-in mic over low-quality AirPods or CarPlay input.
Models and providers
Adding agents that are not built-in providers, such as DeepSeek Harness run as a local ACP command.
Keeping work and personal accounts side by side and switching between them from the thread's model picker.
Adjusting reasoning effort, for example through fx's Effort control, Kiro's per-model Reasoning setting, or ACP agent options in the composer's model picker.
Picking up new logins, agents, or models. Periodic health checks do not reload the catalog, so a manual refresh is needed.
Switching to a text generator, a model API that only writes text. It is usually faster and cheaper than starting a coding-agent session for each request.
Checking remaining quota and upcoming reset times before starting large work.
Summarizing large history when you return to a thread after Claude's prompt cache has expired, instead of resending the full history.
Signing in to Cursor, Codex, or ACP agents while connected to a remote environment. For ACP agents, the sign-in page opens on your device while the agent runs on the environment.
Using OpenRouter models through a Claude instance, or adding a Muse model your account can actually use.
Permissions and secrets
Making every new thread start in the mode you trust, instead of copying the mode of the thread you were viewing.
Approving or rejecting agent requests in the conversation so the agent can continue, and answering subagent approval requests in the parent conversation without switching threads.
Pointing credential fields such as provider environment variables and credentials, source control tokens, usage limit keys, and device host tokens at a 1Password secret reference instead of saving the credential itself.
Giving an agent a credential safely; most requests keep the value from the agent and pass it straight to the tool that needs it.
Using an existing Connect deployment, for example on a home server or in Kubernetes, as the secret source for a headless T3 Code server.
Stopping product events and relay tracing from your server and desktop app.
Use it to cap the broadest permission mode, restrict which providers may run, or lock settings such as agent browser access across an organization.
Use it to confirm that a deployed change is being enforced. Locked controls in Settings show Managed by your organization.
Use it to keep a machine's server off the network. Or set connect to false to turn off only T3 Connect and keep other remote access.
Use it to lift restrictions or roll back a policy. Users get back the settings they had chosen.
Use it to recover from unreadable JSON, unknown settings, wrong value types, unsupported keys, or a machine with both a native policy and a file.
Use it so T3 Code can read secrets on a headless host, a VM, or a background service you reach remotely.
Use it to confirm the token works before relying on references. The test checks your shell, not the server's environment. To confirm the server itself can read the secret, use the reference in a setting and then use that feature.
Use it to reference items whose titles can't be written as a reference, for example an item titled Service Token: CI.
Use it to keep sensitive vaults out of agents' reach, while still letting T3 Code resolve the secrets it needs.
Use them to stop users, agents, and scheduled tasks from running threads above an approved mode, or on providers you have not allowed.
Source control and worktrees
Letting T3 Code clone, publish, create pull requests and review changes on your code host.
Making agent worktree branches match your team's naming conventions, such as issue IDs or namespaces.
Keeping worktrees on another drive or in a folder you prefer.
Speeding up worktree creation in large repositories or repositories with many nested submodules.
Creating worktrees faster on a Mac, and getting a fast first git status in them.
Making sure new work starts from the latest default branch.
Turning an agent's work into a reviewable pull request without leaving the thread.
Letting the agent react to CI failures and review feedback on its own pull request.
Reclaiming disk space, since worktrees add up fast; one user cleared 500 GB.
Laying out clones like ghq, under host, owner, and repo folders in the add-project base directory.
Reading back earlier command output and copying everything the terminal retained.
Checking what an agent changed before you commit, push, or open a pull request, or reviewing a single local commit or a linked commit from a code host.
Keeping your own fish feature flags or re-enabling terminal queries in T3 Code terminals.
Outside agents and subagents
Letting Claude Code or Codex in your own terminal, ChatGPT, Poke, Claude, or an MCP-capable bot work with your T3 Code environment without starting it from T3 Code.
Giving the right address to an agent you configure by hand, such as Claude Code, Codex, ChatGPT, or another OAuth-capable MCP client.
Granting a specific agent access to your environment once, with a level you choose. A sign-in lasts 30 days.
Limiting what an outside agent can do. A thread it starts or steers cannot run with more permissions than the mode you chose.
Letting Claude Code or Codex in your own terminal read projects and threads, start and message threads, and check which providers and models are available.
Working with your environment from ChatGPT or from bots running in the cloud.
Cutting off an outside agent's access to your environment immediately.
Browser and previews
Reusing existing logins in the preview browser without signing in to each site again.
Importing from Safari on macOS, which needs Full Disk Access.
Knowing in advance which source browsers will work before you try an import.
Sending links straight into T3 Code so you can ask an agent about the page right away.
Predicting whether a link joins your current thread or starts a fresh one.
Previewing local HTML files and picking up shared prompts as drafts.
Remote access and devices
Reaching T3 Code running on another machine by signing in to the same T3 Connect account on your other device and choosing the environment.
Keeping someone who takes over your T3 Connect account from reaching your machine from a new device without your approval, and reviewing or revoking approved devices.
Using the LAN at home after pairing once through T3 Connect: while connected through T3 Connect or a paired address, T3 Code learns the machine's current LAN and Tailscale addresses and adds them as routes.
Working on a Linux host or Apple Silicon Mac over SSH, where T3 Code starts or reuses a server and opens the port forward for you; the first launch downloads T3 Code's server to the host.
Getting alerts when an agent finishes, fails, needs approval, or asks for input, and following work without opening the app; tap a notification to open its thread. On iOS you can also add the Agent Activity widget.
Fix a host that looks online but whose tunnel will not start because of a Tailscale exit node, a VPN, or a strict firewall.
Rule out captive Wi-Fi, VPNs, ad blockers, DNS filters, and iPhone Connectivity Assist as the cause.
Give maintainers the details they need, including copied diagnostics, host status, trace IDs, and the tunnel log.
Share a thread from the machine that runs it, and pause all of its guest links at once by turning the setting off.
Letting an agent test mobile work while you watch; an agent-opened device floats over the chat, and in the mobile app you tap the device button above the composer.
Make devices from a Simlock pool available in Settings → Devices.
Appearance and shortcuts
On web and desktop, switch themes without leaving chat with the command palette's Change theme command. You can also open the theme picker directly with Cmd+Option+A on macOS or Ctrl+Alt+A on Windows/Linux. Use Change appearance in the command palette to choose System, Light, or Dark.
Match the mobile app to your device, and use Material You Layout on Android to change shapes, spacing, and controls independently of the selected theme.
Build or share your own look. The editor's color picker lets you click an area of the app to find the color to change, and you can export the theme as JSON.
Set a shared default theme for connected clients, or follow a published palette that updates as the server changes it.
Restyle by request, such as making the chat header taller or setting a wallpaper image as the background.
Read formulas cleanly, including in tables, while plain dollar amounts like $5 stay text. Keep Git branch and worktree controls in view throughout a thread.
Free conflicting shortcuts, assign keys to unbound commands such as thread.stop, or write conditional rules with when expressions.
Send multiline prompts your way, steer or queue a message for one send, start a message in the background, and change models without the mouse.
Browse threads by keyboard: Tab moves through each thread and its buttons, F2 renames the focused thread, and Shift+F10 or the Context Menu key opens the focused thread's menu.
Locate earlier messages, changes in large diffs, or text on a previewed page without scrolling.
Getting your bearings by ear right after you switch back to T3 Code from another app.
Built from the T3 Code docs as of Nov 18, releases through Nov 20 and 50 of Theo’s videos.
Unofficial notes. Not affiliated with Theo or t3.gg.