▶ YouTube Summaries

← All summaries
t3dotgg#▶ Watch

T3 Code tour: multi-harness agents, remote machines, inbox sidebar

2026-10-29 Thu ⏱ 41 min t3dotgg

Theo's first dedicated walkthrough of T3 Code, the open-source desktop/web/mobile app for running coding agents with your own subscriptions. It began as a worse open-source clone of the Codex app; since mid-2026 it has grown to roughly 550k users and 140k weekly actives. Three pillars define it: every harness and sub in one place with cross-harness orchestration, first-class remote control of a fleet of machines, and a sidebar designed as an inbox so parallel work doesn't overwhelm you. The rest of the video is a rapid list of features plus an honest list of downsides.

Origin

  • Theo and Julius loved the OpenAI Codex app but could not fix its performance because it is closed source; the Claude Code desktop app was not good enough either.
  • For months T3 Code was "the open-source worse alternative", built part-time by three people. Theo used Codex more than his own app.
  • Around July 2026 the work on sidebar, remote control and subscription handling made it feel unique rather than a clone.

Pillar 1: every harness, model and sub

  • Built in: Codex, Claude, Cursor, Grok, OpenCode, Antigravity, Muse Code, Pi. Any ACP-compatible harness from the ACP registry (Amp, Cline, Copilot, …) or added manually.
  • Multiple instances of one provider with separate home dirs, API keys and env vars, for rotating accounts.
  • Uses whatever CLIs you already have installed and signed in; existing threads and projects get imported.
  • Orchestrator v2 (an MCP): a thread can spawn sub-agents on different harnesses/models in a worktree, spawn threads in other repos, and threads can find and message each other. External tools (e.g. a chat bot over iMessage) can drive T3 Code through the same MCP.
  • Mid-thread harness switching: hand context from one model/harness to another (e.g. Codex → Opus 5.5).

Pillar 2: remote control

  • Everything goes through a websocket server. npx t3 serve hosts T3 Code on any machine with a minimal web UI; works well over Tailscale.
  • Each new thread picks which machine it runs on; optional load balancing routes to the least-loaded box using CPU/memory diagnostics.
  • T3 Connect: free tunnelling for up to three devices, set up with npx t3 connect. Terminal over T3 Connect is more stable than SSH on mobile.
  • Device hub: iOS/Android simulators and a browser run on a remote machine and stream to the desktop or phone app; Macs can be dedicated as simulator hosts for Linux workers.
  • The mobile app is open source and, in Theo's view, far better than the Claude Code or Codex mobile apps.
  • Threads can be exposed as webhooks (PR opened, Stripe event, plain curl), with an optional 24-hour hold-and-replay.
  • Multiple connection routes per machine (Tailscale, T3 Connect, LAN) with granular permissions; copy an MCP URL per machine.
  • One update button updates the app and the provider CLIs across all machines. "Open in VS Code" opens remote projects over SSH automatically. Projects can be ported (cloned, optionally with env files) to another machine.

Pillar 3: sidebar as an inbox

  • Opening a thread should end with closing it: merged, answered, or no longer needed.
  • Threads are settled, not archived (undo with Cmd-Z), or snoozed to come back later. Threads waiting for input jump to the top.
  • An optional working section hides threads while they run; pinned threads stay visible.
  • Cmd-Enter sends a prompt in the background so you can fire off several in a row.
  • PRs opened by a thread get linked to it; when the PR merges, the thread settles itself (configurable per thread). Idle threads auto-settle after 3 days by default.
  • Multi-select threads to settle, snooze or regenerate titles. Swipe gestures on mobile.
  • The aim is "spatial awareness": knowing where your work is without everything in your face.

Other features

  • Fully open source, including the T3 Connect backend.
  • Agents can render interactive HTML visualizations inline.
  • Built-in PR viewer and management (desktop and mobile), sortable by merge readiness; low-cost PR monitoring that uses far less of your GitHub rate limit than ad-hoc agent polling. Supports GitHub, GitLab, Bitbucket, Azure DevOps, Forgejo and others.
  • Usage tab shows token spend and remaining limits per account, even for runs outside T3 Code. Prompts can be scheduled to fire when rate limits reset.
  • Scheduled tasks: a thread can re-run on a cron-like schedule.
  • Steer vs. queue for follow-up messages, with a configurable default.
  • Speed: switching threads moves about 100x less data than Codex, so it works on bad airplane Wi-Fi.
  • Drafts and stash (Cmd-S) for prompts in progress; drag a thread into another to share context; annotate text or browser content and send it with the next prompt.
  • Attached files are written to disk so agents can use the actual bytes (e.g. attach a screenshot to a PR).
  • Snapshots: press both Shift keys to capture screen and accessibility context into a new thread.
  • Themes (including VS Code import), keybindings and settings scoped per project or per machine.
  • Strong Linux support (almost half of events); ships with Omarchy.
  • npx t3 triage hands diagnostics to Claude Code or Codex to fix local setup issues and file an issue. Theo says it resolves over 80% of problems.
  • Web version at app.t3.codes.

Downsides

  • Bring your own compute and your own subscriptions. T3 Code sells no AI access (that is T3 Chat).
  • T3 Connect is limited to three devices for now; beyond that you need Tailscale or similar.
  • No built-in computer use yet (works only if the harness provides it).
  • You will burn more tokens because you run far more threads in parallel, and it may wreck your sleep.

Takeaway

T3 Code positions itself as built by engineers for engineers, while ChatGPT and Claude push toward general "work" apps. Stable gets all the features shown; a nightly channel gets them sooner.