Happy vs T3 Code
Happy and T3 Code are both free, MIT-licensed apps for running Claude, Codex, and Grok agents on the subscriptions you already pay for, from a desktop app or a phone, and both run other coding agents over the Agent Client Protocol (ACP). Pick T3 Code for a web client, tools for app work, and native Cursor, Pi, and Muse Code integrations. Pick Happy, the T3 Code alternative for teams, for a self-hosted server where everyone signs in as themselves, one harness that runs Claude, Codex, and Grok under the same sandbox, and review on by default.
Side by side
Facts about T3 Code come from its source and docs, checked on October 27, 2026. It ships often; check there for the current state.
| Happy | T3 Code | |
|---|---|---|
| Agents | Claude, Codex, Grok; any ACP agent (beta) | Claude, Codex, Grok, Cursor, Pi, Muse Code; any ACP agent |
| Apps | Desktop, iOS, Android | Desktop, web, iOS, Android |
| Agent loop | Happy's own, or the vendor's (beta) | Each vendor's own |
| Sandbox | One OS sandbox | Each provider's own |
| New sessions start in | Auto: sandbox plus review | Full access |
| Model switch keeps | The transcript | Short chats whole, long ones a selection |
| Team server, per-member sign-in | ✓ | ✗ |
| Scripting | REST API | MCP, t3 thread |
Both run subagents on other providers, steer a running turn, and pick threads back up after a restart.
Where T3 Code is better
- Native integrations. Cursor runs through its SDK, Pi through its RPC mode, and Muse Code through its SDK. Happy reaches them only over ACP.
- A web client. app.t3.codes reaches a machine over HTTPS or T3 Connect from any browser.
- Queued messages you can edit. They are saved on the server, and you can edit them, remove them, or send one as a steer.
- Tools for app work. A live iOS Simulator or Android Emulator next to the thread, browser tabs you share with the agent, and pull requests on GitHub, GitLab, Bitbucket, and four more hosts.
- Guest links (beta). Over T3 Connect, let someone follow one thread in a browser for 24 hours; their messages wait for you to send them.
Run one agent server for your team
Happy Agent runs in team mode on a machine you control. Each member signs in as themselves, with Happy Social (through WorkOS) or your own identity provider. Everyone works in the same projects, sessions, and tasks, under the same sandbox and review rules, and the agent is told who is speaking. Each member pairs their own phone. Phones go through Happy's relay, end-to-end encrypted, so they can read chats even while the server is offline.
T3 Code has no per-member sign-in: you pair devices to a machine with scoped access, and a guest link shares one thread.
Team setup is multi-step today. We haven't built a remove-member action yet, and members' work runs on the server's Claude, Codex, and Grok accounts. See Multiplayer & Teams.
One harness for every model
By default Happy Agent runs the agent loop itself. Claude gets Claude Code's prompts and tools, GPT models get Codex's, and Grok gets Grok Build's. The permission model, the OS sandbox, and the session store are shared. Switch models in the middle of a conversation and the new model continues from the same transcript, with each provider's data kept as it was sent.

Happy can also run a vendor's own agent (beta): Claude Code through the Agent SDK, Codex through codex app-server, or any agent that speaks ACP. Most coding agents ship an ACP command or adapter: Cursor has agent acp, fx has fx acp, Pi has pi-acp, Muse Code has an unofficial adapter, and Antigravity and OpenCode have their own. There is no settings screen to click through. Ask your agent, and it adds the harness to happy.toml for you. The conversation stays in Happy, and Happy's permission mode is mapped onto the agent's own.
T3 Code always runs each vendor's agent, over ACP for Antigravity, Grok Build, fx, and Registry agents. When you switch providers, short conversations transfer whole; longer ones become a budgeted selection, 16,000 tokens by default, without the old provider's reasoning or tool state. The agent can look up the rest with a tool.
Auto by default, not Full access
New Happy sessions start in Auto. Commands run in the OS sandbox, and only an action that would cross it goes to review by a separate agent. Reads and workspace edits never wait on review, so it stays out of the way, and we see no reason to start without it. See Permissions & Sandbox.
New T3 Code threads start in Full access: commands and edits run without approval prompts. You can choose another default in its settings. Its Auto uses each provider's own automatic review, and falls back to asking on providers without one, such as OpenCode and Antigravity.
Updates in the middle of a turn
Install a Happy Agent update while an agent is working: the running step finishes and is saved, the new version starts, and the same turn continues. The window doesn't reload. See Updates Mid-Turn.
T3 Code resumes supported threads after an update or restart by default. Its docs note that a server update can interrupt running agents and terminal commands first.
Drive it from a script
Happy Agent's REST API sends messages, steers a running turn, stops it, and answers an agent's questions, on your machine or a remote one. See Drive Happy Agent over its REST API. T3 Code offers an MCP server for outside agents and a t3 thread command on the host.
Contributing
Both are MIT licensed. We take pull requests, bug fixes first. T3 Code's README, as of October 27, 2026: "We are (mostly) not accepting contributions yet. Small fixes may be considered. Big features will not be."
Which one
| You want | Use |
|---|---|
| Cursor, Pi, or Muse Code without ACP | T3 Code |
| A web client for your machine | T3 Code |
| A team on one agent server, each signed in | Happy |
| One sandbox and permission model for every model | Happy |
| Review before anything leaves the sandbox, by default | Happy |
To start with Happy, see the Quick Start or download it.