跪拜 Guibai
← All articles
Backend · Developer · Artificial Intelligence

T3 Code Unifies Every Coding Agent Under One Control Panel

By cxuanAI ·
Read original on juejin.cn ↗ Google Translate ↗ Alt translation

For developers juggling several AI coding clients, T3 Code removes the need to switch windows and import projects repeatedly. It standardizes agent interoperability through ACP, which means the same control panel can drive code agents from any vendor, and it lets teams parallelize work across different models without file conflicts.

Summary

T3 Code, built by Theo Browne's Ping Labs and released under MIT, positions itself as an "Agent Harness Control Surface." Instead of aggregating model APIs or acting as a token reseller, it takes over the official CLIs and agents already installed and logged in on your machine. Claude Code, Codex, Qoder, Kimi, DeepSeek, and others keep running on their own harnesses, so you avoid the compatibility bugs that come from shimming models into foreign harnesses.

The project brings cross-agent orchestration, Git Worktree isolation, and remote mobile control via Tailscale or its own T3 Connect tunnel. It scans local project paths and imports existing threads from other coding agents. A built-in internal /mcp endpoint lets a main thread delegate subtasks to different agents, each in its own worktree, and collect results back for review and merging. The server runs on Bun and talks to the frontend over WebSocket and JSON-RPC, with a React/Vite web UI and Electron desktop shell.

A notable workflow test showed multiple agents fixing separate bugs in parallel, a third agent verifying fixes, and a final agent reviewing and merging, all within one main thread. T3 Code also handles Git operations like diff, commit, push, and PR creation from the right panel. For Gemini users, T3 Code uses Google's official antigravity-acp server through the ACP protocol, making it the first stable way to use Gemini with personal OAuth without getting banned by reverse proxies.

Takeaways
— T3 Code is an open-source project by Ping Labs (Theo Browne's team), licensed under MIT.
— It acts as a control surface for multiple coding agents, using the original agent harnesses instead of replacing them.
— It supports the ACP (Agent Client Protocol) and the official ACP Registry, so tools like Qoder CLI, Kimi Code, DeepSeek Harness, goose, and Junie can plug in directly.
— Installation is available via brew (brew install --cask t3-code) or npx (npx t3@latest).
— It provides desktop clients for macOS, Windows, Linux, plus web, iOS, and Android, with remote access via LAN, Tailscale, or the built-in T3 Connect tunnel.
— Each task thread gets an isolated Git Worktree, allowing multiple agents to modify the same project without overriding each other.
— The main agent can delegate subtasks to different agents through an internal /mcp endpoint, using ACP to run CLI sub-processes like Qoder in the background.
— Google's Antigravity is an ACP Server on the public registry, so T3 Code can use Gemini via official antigravity-acp runtime with personal OAuth, which is compliant and avoids reverse-proxy bans.
— The backend runs on Bun and communicates with the frontend over WebSocket and a lightweight JSON-RPC protocol.
— The test workflow used GPT-6.2 Sol to introduce a bug, Kimi and Qwen to fix two bugs, DeepSeek-V4.1 Flash to verify, and GPT-6.2 Sol to review and merge.
— The server exposes internal /mcp endpoints for subtask delegation, thread read/write, scheduled tasks, and preview control.
— The architecture separates frontend control surface, local core engine, internal /mcp service, and the dual-track execution of native harnesses plus ACP tools.
Conclusions

The BYOS approach shifts value from model access to orchestration, so T3 Code monetizes through convenience rather than taking a cut on API tokens.

Because T3 Code delegates to official CLIs, it sidesteps the fragile model-harness compatibility layer that many relay clients suffer from.

Standardizing on ACP means Google, OpenAI, and others can make their agents interoperable, which may reduce vendor lock-in for coding agents.

Git Worktree isolation turns parallel AI development into a safe operation, making it practical to run several agents on the same repository without branch conflicts.

The internal /mcp delegation endpoint turns any agent into a worker in a larger orchestrated pipeline, which is a pattern likely to spread beyond coding agents.

Concepts & terms
Agent Harness Control Surface
A control plane that manages multiple coding agents without replacing their underlying execution engines. It provides a unified interface to launch, monitor, and coordinate different AI agents, each running with its original configuration.
ACP (Agent Client Protocol)
An open protocol that standardizes how client applications communicate with AI agents. It allows tools like Qoder CLI, Kimi Code, or DeepSeek Harness to be called programmatically from another application, enabling interoperability.
Git Worktree
A Git feature that lets you check out multiple working trees from the same repository in separate directories. T3 Code uses this to give each agent thread its own isolated branch, so parallel tasks don't conflict with each other.
BYOS (Bring Your Own Subscription)
A model where the tool doesn't sell API access or tokens. Instead, the user brings their own already-paid subscriptions, and the tool just orchestrates them, keeping billing and accounts with the original providers.
From the discussion
Featured comments
用户55568270537

The pain point of switching back and forth is spot on; having each Agent with its own client is indeed fragmented. What impressed me most is the positioning of not selling tokens and only being the control layer; BYOS is quite restrained. The feature of remotely checking progress on your phone also looks pretty practical.

See top comments, translated →
Source: juejin.cn ↗ Google Translate ↗ Backup ↗