T3 Code: An Open-Source GUI for Managing AI Coding Agents
T3 Code is a desktop application built by Theo and the T3 team that acts as a front-end for AI coding agents. It is not an AI model itself. Instead, it connects to existing services like OpenAI's Codex or Anthropic's Claude Code using your own API keys, providing a graphical interface for interactions that would otherwise happen in a terminal or a cloud-hosted web app.
The application is free and open-source. There are no subscription fees layered on top of your existing API costs. This "bring your own key" (BYOK) model means you pay only for the model usage your own account accrues.
The unified interface problem
Developers working with multiple AI agents often find themselves splitting their workflow across several tools. A terminal-based agent may be preferable for one task while a web app works better for another. Switching between them means losing conversation context and project state.
T3 Code is designed to sit in the middle of this fragmentation. Its model selector supports multiple providers from a single interface.
Codex is the current primary integration. Claude Code support has a completed pull request awaiting approval. Cursor, OpenCode, and Gemini are planned. The goal is that switching the underlying model becomes a single dropdown selection rather than a tool switch, with project context and thread history preserved across providers.
Interface layout
The application uses a three-panel layout.
The left sidebar lists local Git repositories that have been added to T3 Code, using project favicons where available for quick identification. The middle panel shows all threads for the selected project, where each thread represents a distinct task or conversation. The right panel is the main interaction area: a chat-style interface that displays prompts, responses, the agent's plan, and real-time tool calls.
Adding a project and configuring an agent
Projects are added by clicking Add project in the sidebar and providing the path to a local Git repository, either by browsing or pasting an absolute path directly. For developers who work primarily in the terminal, pasting the result of pwd avoids navigating through folder trees.
Before sending a prompt, several configuration options are available in the interaction panel:
- Model selector: chooses the underlying agent or provider.
- Reasoning level: controls how much planning the agent does before acting. Higher levels produce more thorough plans at the cost of speed.
- Chat vs. plan mode: chat mode is suited for quick iterative questions. Plan mode instructs the agent to inspect the codebase and generate a full step-by-step plan before writing any code, which is more appropriate for larger or more complex tasks.
- Full access vs. supervised mode: full access lets the agent read and write files without prompting. Supervised mode requires approval for each action.
Git worktrees for isolated agent work
T3 Code has native support for git worktree, which allows a repository to have multiple checked-out working directories on different branches simultaneously.
When a task is started in a new worktree, the agent performs all its modifications in a separate directory. The primary working directory is not touched. This means an agent can make extensive changes to a feature branch while you continue working on the main branch in parallel. Once the agent finishes, the changes can be reviewed with git diff, tested, and merged only when they are confirmed correct. T3 Code handles the underlying Git commands automatically when the worktree option is selected.
Commit, push, and pull request in one action
After an agent completes its task, a set of controls in the top-right of the interface handles the entire commit-to-PR workflow.
The Commit button opens a modal showing a summary of changed files with line counts added and removed. A commit message can be written manually, or left blank for the application to auto-generate one based on the agent's work.
The Commit, Push, & Create PR button chains all of this together in a single click: it generates the commit message if needed, commits the changes, pushes the branch to the remote, and opens the browser directly to the new pull request page on GitHub with the title and body pre-filled.
The thread sidebar icon updates to reflect the pull request status (open, merged, or closed), so the state of each task is visible without leaving the application.
Diff viewer and integrated terminal
The built-in diff viewer lets you review the agent's changes turn by turn, seeing exactly what was modified at each step of the plan. Both unified and split views are available.
An integrated terminal is also available within the application for running linters, formatters, build tools, or any other shell commands without switching to a separate terminal window.
Custom quick actions
Quick actions are project-scoped buttons that execute a shell command on demand. They are created by clicking the play icon in the thread toolbar and selecting Add Action. Each action takes a name, an optional keybinding, and a command such as npm run dev. Actions can be configured to run automatically when a new worktree is created, which is useful for setup commands like npm install.
Performance
Desktop applications built on Electron are often criticized for high memory usage and sluggish UI. T3 Code is noticeably responsive: switching between projects and long threads is instant, and running multiple agents in parallel does not degrade the interface. This is a meaningful practical difference for tools that stay open throughout a working session.
Planned features
T3 Code is under active development. A few capabilities that would extend its usefulness are currently being discussed or built:
- Skills support: a UI for managing reusable complex prompts, invokable inline with syntax like
@skill-name, similar to the skills system in the original Codex app. - Headless/remote mode: the ability to run the T3 Code backend on a remote server and connect to it with the local client, useful for offloading resource-intensive tasks and enabling collaborative sessions.
- CLI integration: a command like
t3code .that opens the current directory as a project in the app, closing the gap between terminal and GUI workflows.
The project is open-source and actively maintained at github.com/t3-oss/t3-code. Feature requests, bug reports, and pull requests are welcome.
Final thoughts
T3 Code occupies a useful gap between terminal-based agents and cloud-hosted coding assistants. The Git worktree integration, one-click PR workflow, and per-turn diff viewer are practical additions that address real friction in AI-assisted development. The BYOK model keeps costs transparent, and the multi-provider roadmap makes it worth following even if your current workflow relies primarily on a single agent.
For developers who already use AI coding assistants regularly and find the context-switching between tools disruptive, T3 Code is a practical improvement worth evaluating.
This work is licensed under a Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License.