T3 Code: An Open-Source AI Companion for Developers

T3 Code was conceived by Theo and Julius as an open-source solution after they encountered limitations with OpenAI's Codex, which became less performant, and were denied access to its source code. They also found Anthropic's Claude Code desktop app to be unsatisfactory. Initially, T3 Code was maintained by a small, part-time team and struggled to compete with the extensive resources of trillion-dollar tech companies.

Despite its early challenges, T3 Code began to evolve. Julius significantly advanced the project, and Theo contributed to UI and UX improvements, implementing features that proprietary labs were not interested in, such as changes to the sidebar, remote controls, and subscription integrations. Around July 2026, these developments shifted T3 Code from a simple alternative to a unique product.

This transformation is reflected in its user growth, with the project nearing 550,000 users and 140,000 weekly active users by mid-2026, a milestone described as "surreal" by Theo.

T3 Code now provides native integrations for a wide range of models and subscription harnesses. These include Codex, Claude, Cursor, Groq, OpenCode, Anti-Gravity, MuseCode, and Pi, all directly built into the application. This expanded support aims to enable users to work with virtually any model or harness they utilize.

Beyond native support, T3 Code incorporates the full Model Context Protocol (MCP) registry, allowing users to add other compatible tools such as AMP, Klein, and Copilot. Users also have the flexibility to manually add other tools if needed. This comprehensive approach means "pretty much everything is now supported."

For users with multiple subscriptions, the platform allows adding multiple instances of the same provider. Each instance can be configured with different home directories, API keys, and environment variables. This feature is particularly useful for managing various accounts and testing early access programs. The system supports advanced workflows, such as reviewing a single codebase simultaneously using four different models like GBD61 medium ultraast, Soul Muse Wonderful Full Gemini 4 Argan, and Sonnet 55.

T3 Code facilitates the concurrent use of multiple AI harnesses and agents within a unified interface. Users can deploy various agents, including Gemini and Sonnet, with configurations seamlessly drawing from existing user subscriptions, eliminating the need for special setups. This capability enables agents to work together efficiently.

The system allows for cross-thread communication and task delegation, enabling a single agent to spin up sub-agents in other harnesses and models. A demonstration showed four sub-agents, each with a different harness, operating within T3 Code, including GBT61 Soul Muse Wonderful Full Gemini 4 Argan and Sonnet 55. These agents use their official harnesses and the user's existing subscriptions without requiring special configuration.

Users can also manage multiple repositories from one thread. For instance, a user can initiate a thread to explore a codebase and find easy wins, and then instruct it to spin up another thread in a different repository with the same request. These threads can communicate directly, even allowing users to copy a thread's ID to enable direct interaction between them.

T3 Code functions as a flexible orchestration layer, rather than imposing a predefined workflow. The Model Context Protocol (MCP) structure, developed by Julius, provides agents with the flexibility to perform a variety of tasks and integrations, allowing users to define how they want to work.

The system supports integration with external tools such as Grockbot, enabling users to manage T3 Code threads entirely from outside the application. For example, one user reportedly managed over 10,000 threads through an instant connection via iMessage, demonstrating the platform's adaptability to external management.

The orchestration layer is designed for ease of use: installing the T3 Code app and having existing harnesses set up automatically imports existing threads and projects, providing immediate access to these capabilities through a user interface that is considered less cumbersome and faster than alternatives.

T3 Code explicitly requires users to provide their own subscriptions for AI access, differentiating it from T3 Chat, which is a closed-source alternative to ChatGPT. T3 Code is positioned as an open-source tool for working with AI agents, models, and harnesses.

T3 Code features built-in support for device simulators, allowing users to run iOS and Android simulators directly within the tool. This functionality is part of the broader device hub capabilities, designed to streamline mobile application development and testing.

The platform also enables users to dedicate specific machines for certain tasks. For instance, Macs can be dedicated for iOS builds, allowing other computers in the network to SSH into them to run development servers for iOS apps, even while the primary work occurs on a Linux box. This feature, developed by Julius, aims to simplify interactions with tools like Xcode and improve the iteration process for the T3 Code mobile app.

Additionally, T3 Code's mobile app is highlighted as a highly functional open-source solution for remote AI code tasks, noted as superior to the mobile apps of Cloud Code or Codeex, which are described as ineffective for remote work.

T3 Connect offers a simplified tunneling service for remote device connectivity, allowing users to connect up to three devices for free. This system aims to eliminate the need for manual VPN or SSH management, particularly for mobile use, by providing a direct and stable connection.

The service enables users to manage remote machines entirely from their desktop or mobile app. For example, by running `npxt3 connect` on a Linux box, users can link it to their T3 Code account and then manage the entire machine remotely without further SSH intervention. This approach is described as a pleasant way to manage AI coding tasks, leading to fewer threads being run on personal computers.

The integration of T3 Connect with an old Linux computer is highlighted as particularly effective for handling tasks like Rust builds remotely, making such processes trivial to manage.

In T3 Code, the sidebar functions as an inbox where the primary objective for every open thread is to 'close' it. Closing a thread can mean various actions, such as merging a code change, providing an answer to a question, or determining that a task is no longer necessary. This approach is designed to encourage users to actively resolve or clear out tasks.

To support this, T3 Code employs a 'settle' function instead of permanent archiving, making actions less final and easily reversible with a command-Z option. Settled threads are moved to a designated section at the bottom of the interface. Users can also 'snooze' threads until a later time, effectively postponing their visibility in the main inbox.

The system also includes visual indicators, like an input indicator, to highlight threads that require user attention or a response, helping users prioritize their tasks within the inbox-like environment.

T3 Code introduces a 'working' section, an optional beta feature that hides threads while they are actively processing. This design choice aims to reduce visual clutter and help users maintain focus, especially those who may find too much concurrent information distracting.

Despite initial concerns that users might dislike this feature, it has gained popularity, with many finding it indispensable. For users who prefer to keep an eye on active processes, threads can be 'pinned' to remain visible even when they are in an active 'working' state, providing a workaround for individual preferences.

The interface supports parallel task execution through a 'command-enter' shortcut, which sends prompts to run in the background. This allows users to initiate multiple tasks quickly without being interrupted by the immediate display of results, further promoting a focused workflow.

This design philosophy aligns with the goal of enabling parallel work without overwhelming users, contrasting with other AI building and coding tools that tend to display everything upfront. Once a background task finishes, its thread reappears at the top of the interface.

T3 Code is fully open source, with its entire repository, including infrastructure backends like T3 Connect, publicly accessible. This transparency allows users to fork the project and run it independently, even connecting their forks via T3 Connect for remote operations. This open nature ensures that all aspects of the project are visible and modifiable by the community.

T3 Connect provides a stable alternative to SSH for remote connections, particularly effective for mobile use. It is described as being more stable than SSH, especially on mobile devices. This enables seamless management of remote terminals and prompts within the T3 Code environment, supporting various remote computing tasks.

The platform also integrates advanced features like full diagnostics for connected machines, allowing users to monitor compute usage across their setup. Additionally, agents within T3 Code can render full HTML visualizations with interactive JavaScript, facilitating the creation of charts, diagrams, and other complex visual outputs directly within the app. Users can also manage GitHub pull requests directly within T3 Code, including merging and other actions, which is also available on the mobile app.

T3 Code is optimized for efficiency, significantly minimizing the data required for operations. Hopping between threads in T3 Code uses "a hundred times less data" than tools like Codeex, which is noted to have become more bloated, while T3 Code has become "trimmer."

This optimization ensures that T3 Code performs well even on poor network conditions, such as 3G mobile connections or low-quality airplane Wi-Fi. The developer notes being able to code effectively on airplane Wi-Fi without experiencing slowdowns, which only become apparent when switching to a browser.

Beyond data efficiency, T3 Code offers features like shift-clicking to select multiple threads for bulk actions (snoozing, settling, regenerating titles) and auto-settling threads that haven't been touched for a configurable period (defaulting to 3 days). Users can also access T3 Code through a web app, connecting to their NPXT3 instances running on other machines without needing to install the desktop application.

T3 Code now supports switching between different AI models, or 'harnesses,' mid-thread while maintaining context. Previously, users were confined to the harness they started with, as threads were tied to specific harnesses. The new system allows the context to be transferred between harnesses, enabling seamless transitions between models like Opus 5.5 without losing the ongoing work or connections to devices.

The platform offers extensive operating system support, including Windows ARM, despite the developer's stated dislike for it due to increased build times and frequent failures. This broad support is maintained due to user feedback, as deprecating any platform tends to draw negative reactions. Linux users represent a significant portion of T3 Code's events, highlighting its strong support and integration within the Linux ecosystem, to the point where it's now built directly into Omari.

Users can also manage their Linux agents from T3 Code on a Mac, control other threads by dragging them into a new thread to provide context, and add annotations to specific parts of conversations or browser previews. The system supports various Git providers like Azure DevOps, GitLab, Bitbucket, Forge Joe, Git Cafe, and GitHub, facilitating integration into diverse development workflows.

T3 Code incorporates built-in PR monitoring flows that are significantly optimized to reduce GitHub API rate limit usage. The developer claims these flows use as little as a hundredth of the API rate limits compared to agent-based solutions that might repeatedly babysit PRs. This makes the monitoring process much more efficient and reliable.

The system also supports multiple connection routes to a given machine, including Tailscale, T3 Connect, or LAN, with a granular permission system for remote connections. Users can expose their MCP server to other machines with a one-click copy URL feature, and threads can be exposed over webhooks.

Webhook integration allows threads to be triggered by specific actions, such as when a PR is opened or an event occurs in Stripe. Users receive a URL to forward events to their environment, or they can opt to have events held and replayed for up to 24 hours for added durability. The ability to bring a custom public URL allows users to fire a curl command to trigger a thread remotely. Additionally, T3 Code automatically uses project favicons as icons for projects, with options to manually choose or generate icons if needed.

T3 Code now offers one-click updates for the application and integrated AI providers like Claude and Codeex across multiple devices simultaneously. This feature simplifies maintaining up-to-date environments, especially for users with numerous machines. For example, the developer stated they update T3 Code and its providers across nine different computers with a single click.

A notable new feature is the 'Open in VS Code' button, which automatically opens projects in VS Code's SSH mode, connecting remotely to the machine where the code resides. This functionality allows users to edit code, environment variables, or markdown files on remote machines directly from their local VS Code instance, even if the project is not physically on their local computer.

The platform also introduced the ability to port projects between machines. Users can clone and set up entire repositories, including copying environment variables, onto any connected machine with a single click. This feature enhances flexibility for developers working across multiple systems, ensuring consistent project setups.

T3 Code provides a unique capability for AI agents to access raw files and images directly. When an image is attached, it's not only passed up as content but also placed on the system where agents can access its raw bytes. This enables agents to perform tasks like "add this image to the PR," a functionality not typically found in other tools that can only see image content.

This direct file access extends to various file types, allowing agents to embed anything. The platform also enables the T3 app to render videos and photos over remote connections, which is useful for tasks such as attaching a video of a code fix to a pull request. The developer noted that unlike the GitHub CLI, T3 Code's agents have a built-in file upload skill to embed images.

Another feature, 'snapshots,' allows users to press both shift keys to capture context from their current application (e.g., a browser) and automatically open a new thread in T3 Code with that information. This includes not just the app's content but also accessibility data from the system, providing richer context for agents. T3 Code also offers superior custom font support compared to other tools.

Theo openly discussed several limitations of T3 Code. Primarily, users must "bring their own compute," meaning T3 Code does not provide a cloud provider, requiring users to set up their own infrastructure for the application to run. This allows users to deploy it in their own cloud environments.

T3 Connect currently has a limit of three connected devices. Users with more than three machines need to rely on alternative connection methods like Tailscale. While future plans include expanding this limit, the current restriction is in place to prevent the open-source project from incurring excessive costs. Another limitation is that computer use is not yet natively built into T3 Code, though it supports external Model Context Protocols (MCPs) from other harnesses like Codeex, which do offer this functionality.

Theo humorously noted a significant "issue" being that the project is "made by Theo," but quickly highlighted that its open-source nature allows users to fork the project if they have objections. The tool is also acknowledged to potentially "ruin your sleep schedule" due to its addictive nature, leading to late-night coding sessions.

Despite these limitations, the open-source model allows for community contributions to address shortcomings, and the core team plans to integrate more features, such as native computer use, in the near future.

T3 Code offers users a choice between a stable version and a nightly version. By default, downloading T3 Code from the official site provides the stable release, which includes all the recently showcased features. This version is intended for users who prioritize reliability and a polished experience.

For those who prefer to access the latest features and are comfortable with potential instability, a "nightly button" is available on the download page. This option provides access to bleeding-edge updates as soon as they are ready, though it comes with the understanding that changes may occur frequently and occasionally lead to temporary breakages during the iteration process. The nightly version guarantees much faster updates for users who want to stay at the forefront of development.