T3 Code: Unified Dashboard for Multiple AI Coding Agents with Mobile Control
T3 Code is an open-source orchestrator that unifies AI coding agents like Claude Code, Codex, and Cursor into a single dashboard with web, desktop, and mobile access, enabling model switching within threads, sub-agent delegation, remote approval via phone, and Git worktree integration, while noting caveats like default permissive permissions and early version status.
What Is T3 Code?
T3 Code is not a harness — harnesses like Claude Code or Codex turn models into agents. Instead, T3 Code acts as a scheduler that collects your already-logged-in CLIs into one control plane. The software is free with no paid tier; you still pay Anthropic, OpenAI, etc. for their subscriptions. T3 Connect (the relay service) is also free, limited to three machines per account (phones and browsers don't count). The only prerequisite: a machine with at least one provider CLI installed and logged in (e.g., claude auth login or codex login). T3 Code does not install or log in for you.
Architecture: Server, Client, Connection Layer
The system comprises three pieces:
Server — runs on the machine holding your code. It manages provider processes, terminals, Git, and project files. The desktop app bundles a server; the CLI command t3 starts an equivalent standalone server.
Client — web, desktop, and mobile views sharing the same connection logic. Clients never touch files or credentials.
Connection layer — T3 Connect for NAT traversal, plus LAN pairing and Tailscale as fallbacks.
The repository contains 31 packages: 10 provider adapters (Claude, Codex, Cursor, Grok, OpenCode, Antigravity, Kiro, Pi, Muse, fx), 7 code-hosting integrations (GitHub, GitLab, Bitbucket, Azure DevOps, Forgejo, Gitea, GitCafe), plus standalone ssh and tailscale packages.
Highlight 1: Model Switching and Sub-Agents in a Single Thread
Each provider speaks through its own adapter, emitting normalized events to an orchestration layer, so a thread isn't bound to a single CLI. Two usage patterns:
Mid-stream switch — start with Claude, hit a snag, hand off to Codex without restarting context.
Delegation — the main thread splits work and spawns multiple sub-agents running in parallel, each possibly using a different model.
The article shows a screenshot where a thread lists "opening Claude, Codex, Muse and Gemini" with four sub-tasks (share button, share link API, link test, docs) completing in 19–25 seconds; the Lineage panel shows "4 running". The benefit for multi-subscription users is avoiding context re-pasting, not just saving keystrokes. The cost: parallel runs consume multiple quotas simultaneously — the author acknowledges quotas drop faster because you're doing more at once, not because of inefficiency. A caveat: provider permission tiers differ (Muse has only full-supervision and fully-open; fx lacks auto-accept-edit; Grok lacks it too). When dispatching across models, the tier is downgraded to the lowest common denominator.
Highlight 2: Remote Access Built into Architecture
Most desktop tools treat remote as an add-on; T3 Code separates server and client from day one, so remote is just a different connection method. One command: t3 up — it asks how this machine should be reachable (T3 Connect, LAN, or both), then installs itself as a background service. t3 status shows current state.
A memorable demo: an agent on the development machine runs pnpm db:migrate; the phone pops up a confirmation card with three buttons — Allow once, Allow for this session, Deny. The author is away, taps once, and the migration proceeds.
Security details: by default, a logged-in account can connect. You can enable new-device approval — a new device shows a code like K7QM-2XFD; an already-connected device must approve within 10 minutes. Headless machines can approve via t3 auth connect-device approve. Code and prompts travel only from your machine to the model vendor; T3 Connect uses Cloudflare tunnels that relay connections without proxying HTTP/WebSocket. Telemetry sends only model name and token counts; disable with T3CODE_TELEMETRY_ENABLED=false.
Limits: free tier = 3 machines (Pro = 10, Premier = 100). No guarantee how long free tier lasts; the FAQ explicitly says they don't mind if you configure Tailscale as a fallback — the author does so himself.
Highlight 3: Git Integrated into Workflow
Each thread can have its own worktree. Changes appear in a diff panel with stats like +71 -2 directly on the commit button. Commit messages and PR descriptions can be generated following the repo's existing conventions. The main view shows "Changes vs origin/main" with the diff stats and a "Commit, push & PR" button.
Seven code-hosting platforms have independent implementations — GitLab, Bitbucket, Azure DevOps, Forgejo, Gitea, GitCafe are not thin wrappers around GitHub's API.
Worth it if your current flow is: agent finishes, you open another window for git diff. The time saved is judgment time, not command time. The cost: disk space. Worktrees are copied per thread; heavy usage grows fast — the FAQ mentions one user reclaiming 500 GB. Auto-cleanup exists but is off by default.
Caveats: Default Permissions, Version 0.1, PR Backlog, Limitations
Default permission is fully open. Four tiers from strict to loose: full supervision, auto-accept edit, auto, fully open. New threads start at fully open — agents can run commands and edit files without asking. This is a product choice for IDE-like fluidity. Mitigation: immediately change Settings → Threads → New threads → Permissions to full supervision, then relax gradually. Teams can enforce a managed policy ceiling. Do not run fully open on machines storing payment or banking data.
Version is still 0.1. Current v0.1.10, previous v0.1.9, nightlies before that. 841 releases indicate rapid iteration; APIs and UX will change.
PR backlog. 24,707 PRs submitted, 11,701 merged (47%); 2,288 open PRs and 1,191 open issues. README admits limited review capacity; small focused fixes merge easier; new features need discussion first. Fine as a user, but contributors need patience.
Does not fix harness limitations. No local model support — FAQ states local models can't do meaningful engineering work. No t3 binary for Intel Mac; only the desktop app works there.
Installation and Configuration (5 Steps)
One-line install: curl -fsSL https://t3.codes/install.sh | sh (Windows: irm https://t3.codes/install.ps1 | iex). Or try without installing via npx t3@latest (requires Node).
Run t3 — starts the service and opens the web UI.
Open Settings → Providers, pick a logged-in CLI; if not detected, set Binary path manually.
For phone access, run t3 up, choose T3 Connect, follow QR-code login.
For persistence: t3 service install enables auto-start on boot.
Configuration Tweaks
New thread default permission — set to full supervision. Location: Settings → Threads → New threads.
Provider binary path — full path to the CLI. Location: Settings → Providers → Binary path.
Telemetry — disable by setting environment variable T3CODE_TELEMETRY_ENABLED=false.
Reclaim disk space on shutdown — run t3 doctor from command line.
Verification and Common Pitfalls
Verify with t3 status (output appears) and t3 status --json (prints config). Web UI should open a thread and hold a normal conversation.
Two real gotchas:
If a t3 server is already running, starting another reports "service exists". Fix: run t3 pair to connect to the existing instance, or use --base-dir to launch an independent instance.
Provider list empty? 90% chance the CLI isn't in the server machine's PATH (not your phone's). Version managers like nvm or asdf often cause this.
Time to first working setup: ~10 minutes, assuming agents are already logged in.
Personal Verdict
The author's benchmark for such tools: can it replace the makeshift setup I currently tolerate? The author currently uses tmux with three panes and checks phone progress via screenshots sent to WeChat. T3 Code replaces both — hence the desire to install.
T3 Code is not in the same layer as Claude Desktop or the Codex App; those are single-vendor clients. T3 Code puts multiple vendors on one desk. The real watershed is the moment you switch models: the thread doesn't restart, context isn't re-pasted.
Shortcomings reiterated: default fully-open permissions, version 0.1, 2,000+ open PRs, free tunneling limited to three machines.
Two profiles should try it: (1) developers already paying for Claude/ChatGPT coding tiers who run two or more agents daily — install, set permissions to full supervision first, then relax. (2) Single-agent, single-machine, no remote needs — skip; T3 Code solves multi-thread/multi-device chaos, not single-thread efficiency.
GitHub repository: https://github.com/pingdotgg/t3code
Signed-in readers can open the original source through BestHub's protected redirect.
This article has been distilled and summarized from source material, then republished for learning and reference. If you believe it infringes your rights, please contactand we will review it promptly.
How this landed with the community
Was this worth your time?
0 Comments
Thoughtful readers leave field notes, pushback, and hard-won operational detail here.
