T3 Code Turns Your Phone Into a Remote for Your Coding Agents. Change Two Defaults First

One free app drives Codex, Claude Code, Cursor and more from your phone. Out of the box, it is also a key to your laptop.

Coding agents multiplied faster than anyone's patience for switching between them. Codex has its CLI. Claude Code has its terminal. Cursor, OpenCode and Antigravity each have their own window, their own login and their own way of asking permission.

So the next layer was predictable: one remote for all of them. And a remote for agents that can run shell commands is, under the friendly interface, a remote for your computer.

That is the honest description of T3 Code, the open-source project from T3 Tools that sat at #2 on Trendshift's daily board this afternoon (November 2). It is useful, it is free, and two of its defaults deserve five minutes of your attention before you scan a QR code with your phone.

What T3 Code is

The README calls it an "agent harness control surface." In plain words: a single app, on iPhone, Android, the web or the desktop, that controls the coding agents already installed on your machine.

It does not bring its own model. It drives the ones you already pay for. You install and log in to at least one provider, and T3 Code takes it from there:

Codex (codex login)
Claude Code (claude auth login)
Cursor's CLI (agent login)
Grok Build, OpenCode, Antigravity and fx, each with its own sign-in

The agents keep running on your computer, or on a server or VPS (a rented virtual server) if that is where you install it. According to the FAQ, your code and prompts travel from that machine to your provider, such as Anthropic or OpenAI, and do not pass through T3's servers. The project is MIT-licensed and free; you keep paying your providers as usual.

As of today it shows about 41,000 stars on its GitHub badge, and several nightly builds tagged v0.1.7-nightly.20261102 (with build suffixes such as .4066) shipped today. The README is blunt about its age: "very early, expect bugs."

How your phone reaches your laptop

There are a few ways in, and the remote access guide covers each:

T3 Connect, the no-router-setup option. Run t3 up and choose it. The FAQ says the relay only sets up the connection and does not proxy your sessions; traffic goes through a Cloudflare tunnel that runs on your machine, an outbound connection that saves you from opening ports on your router. It is free for now, for up to three machines per account.
Direct pairing over your home network or a private network, with t3 up --lan. You get a one-time link or QR code to scan.
Tailscale, the private-network service many developers already use, with t3 serve --tailscale-serve.
SSH, the standard encrypted remote login, managed from the desktop app.

The app picks whichever route answers fastest. That is convenient. It also means your laptop can be reachable from several directions at once, and each one deserves to be set up on purpose.

The two defaults to change

Here is the part most first-run guides skip.

Default one: new threads start in Full access. T3 Code has four permission modes. Supervised asks before commands and file changes. Auto-accept edits approves file edits and may still ask about other actions. Auto uses the provider's own automatic review for routine actions and asks you about the rest. Full access runs commands and edits without asking. The docs say new threads start in Full access.

There is one softening detail. With Claude as the provider, Full access can still prompt if your Claude settings disable bypass permissions or require approval for a specific action, and Claude's dangerous-removal checks still reject commands like a risky delete. The docs note a provider-specific exception for Antigravity too, so check what Full access means for the agent you actually use.

Default two: a pairing link without --scope grants everything. The remote access guide says so in plain terms: without --scope, pairing grants every permission, "including terminal and orchestration operation." The paired device can open a terminal on your machine and run agents, not just watch them. The same guide tells you to treat pairing URLs and codes as passwords and keep them out of screenshots, logs and bug reports. Pairing also lasts: it grants access for future connections until you revoke it.

Put those two together and picture a QR code on screen during a livestream, or a pairing URL pasted into a support thread. The person who grabs it gets something close to a remote shell on your laptop.

That is the idea worth keeping from this tool, and from every tool like it. A phone remote for your agents is a remote shell with a nicer interface. Set it up the way you would set up SSH.

Put this into practice

Plan on about fifteen minutes.

1. Try it once before installing anything

npx t3@latest

This runs T3 Code one time and opens the local web app. If you like it, the permanent install on macOS or Linux is:

curl -fsSL https://t3.codes/install.sh | sh
t3

Read the install script before piping it to your shell, as you would with any curl | sh installer. On a Mac, brew install --cask t3-code gets you the desktop app instead.

2. Change the default permission mode first

Before you start a single thread, go to Settings, then Threads, then New threads, then Permissions, and pick Supervised or Auto. You can loosen it later for a project you trust, since projects can override the default.

What to expect: Supervised will feel slow on day one. That is the point. Watch which approvals you keep clicking, then move only those to automatic.

3. Pair your phone with a scope

Instead of a bare pairing link, give the phone only what it needs. The docs show a read-only example:

npx t3 pair --scope orchestration:read --scope relay:read

That lets a device follow along without running terminals or managing access. It is not file access to your machine either; it covers agent activity and the connection. Run t3 auth --help to see your access options and pick the smallest set of scopes that covers how you actually use your phone. If you want to send prompts from the train, you will need more than read access; just leave terminal control off unless you mean it.

4. Clean up access you forgot about

Revoke pairing links you never used, and device sessions for phones or laptops you no longer trust. t3 auth --help lists the commands.
For SSH environments, t3 auth session list and t3 auth session revoke <id> remove old sessions.
Run t3 connect unlink to stop exposing a machine through T3 Connect.

5. Turn off what you don't need

T3 Code sends basic usage events (provider, model, token totals, not prompts or file contents, per the FAQ). Set T3CODE_TELEMETRY_ENABLED=false if you would rather it didn't. And if you try the beta feature that lets agents on one machine work on another, read the access level carefully: the docs say full access lets linked agents act like that machine's own agents. Links expire after 30 days.

Which mode for which project

The four modes are easier to choose once you think about what a mistake would cost in each project, not how much you trust the agent.

A throwaway prototype or a scratch folder. Auto-accept edits is fine. The worst case is a messy directory you delete. Keep commands on approval, because commands can reach outside the folder.

Your day job's main repository. Auto, if your provider supports it, or Supervised. This is where an agent can push a branch, run a migration script or install a package you never reviewed. Let the provider's review handle the routine reads and test runs, and look at the rest yourself.

Anything with production credentials in its environment. Supervised, every time. If a project folder holds deploy keys, cloud tokens or a .env file with live secrets, an unattended agent there is one bad guess away from a real incident. Better still, keep those secrets out of the folder the agent works in.

A dedicated VPS you set up only for agents. This is the one place Full access makes sense, because the machine itself is the boundary. If the agent wrecks it, you rebuild it from a script. That is also why many people run agent tools on a separate box in the first place.

A useful rule of thumb: the further you are from your keyboard, the stricter the mode should be. Approving a command at your desk with the full diff on a big monitor is one thing. Approving it from a phone in a grocery line is another. T3 Code's whole pitch is the grocery line, so set the defaults for that version of you.

Honest limitations

It is early software. The README says to expect bugs, and the maintainers are mostly not accepting large contributions yet, so a missing feature may stay missing for a while.

Windows is a second-class host for now: no background service, so t3 serve stays open in a terminal.

T3 Connect is free "for now." The FAQ says the team may change course if costs become unacceptable, and linking more than three machines ties into paid T3 Chat plans (10 machines on Pro, 100 on Premier).

Your provider's rules still apply. The FAQ claims T3 Code fits the terms of service for Claude Code, Codex and Google Antigravity, and also warns against juggling many accounts inside it because of ban risk. That is the project's own reading of other companies' terms, so check them yourself if you depend on a subscription for work.

And no setting here fixes a bad agent decision. Supervised mode moves the decision back to you, on a small phone screen, possibly while you are half paying attention. Read before you tap.

The remote is only as safe as its defaults

T3 Code is a good idea done quickly: one place to run every agent you pay for, from wherever you are. The risk does not come from the app itself. It comes from defaults chosen for a smooth first run.

Change the permission mode before your first thread. Scope your first pairing. Then decide, on purpose, how much of your laptop you want reachable from your pocket.

Sources: pingdotgg/t3code README; T3 Code permission modes; T3 Code remote access guide; T3 Code FAQ; GitHub releases; Trendshift daily board (read about 15:13 ET). Docs read on the main branch on November 2, 2026, alongside the v0.1.7-nightly.20261102 builds.