Theo Browne runs Ping and makes T3 Chat, and before that he created the T3 Stack. He also posts several YouTube videos a week, many of them about the tools he codes with, so few developers have documented their switches as publicly. He gives one warning himself: he invests in many of these companies and takes sponsors, and asks viewers to “account for all sorts of biases.”
A boring setup on purpose
In September 2022 he showed his environment to make a point. He had once spent a lot of effort recreating an i3 setup on his Mac, then watched his first manager at Twitch outpace him with nothing but Cmd+Tab. Now he used the stock macOS Terminal, “I don’t use iTerm,” with tmux because “I grew up on tmux,” and VS Code with little more than language extensions: “very boring by design.”
He had also just changed his mind about AI. In July 2022 he admitted “I was straight up comically wrong about copilot”: after more than a month with GitHub Copilot, its suggestions made him faster. By May 2024 he showed his VS Code with “copilot’s off because I’ve moved to Supermaven,” a faster autocomplete he had also invested in.
Cursor
Cursor took him three tries. In September 2024 he said he had invested early but “did not actually like using it and I churned from the product twice.” This time it stuck. In April 2025 he listed it among his daily tools: “Cursor’s become my editor of choice,” and “for the first time in years, the combination of Cursor plus Claude made me feel significantly more productive.”
His December 2025 workflow video opens with “I’m still on Cursor.” He worked in its agent view, planned with Claude Opus, “the model I’m using for most of my code recently,” and ran the same plan on several models at once in separate git worktrees. He dictated prompts with Wispr Flow, scaffolded projects with Bun, used no MCP servers, and reviewed pull requests on GitHub with bots such as CodeRabbit and Greptile. He also called the GitHub CLI “legitimately one of the worst pieces of software I’ve ever used.”
T3 Code
The Codex app changed that. Launching T3 Code in March 2026, he said that working in projects with parallel threads “changed how I write code” and “turned my editor back into an editor.” He then outgrew the Codex app: its performance, its worktrees and “the lock in on just codex models.” So he and Julius Marminge built T3 Code, a free, open source desktop and web app that runs other companies’ agents. At launch it needed a Codex subscription, and even on launch day he left it “to go work in Claude Code in the terminal quickly” for UI changes.
pingdotgg/t3code45.8k- apps/
- assets/
- docs/
- evals/
- infra/
- native/
- oxlint-plugin-t3code/
- packages/
- packaging/
- patches/
- LICENSE
- README.md
By May 2026 he had “entirely changed my workflow,” he said: “I really struggled to use anything other than GPT 5.5” and had “pretty much entirely stopped using Claude models.” He ran Codex’s harness through T3 Code, opened an editor mostly “to edit environment variables now,” and reached his T3 Code over the network from the Helium browser, with a phone app on the way. He was firm about terminals: “I really really don’t like doing agent coding in terminals anymore,” and the effort people put into running Claude Code over SSH and tmux was something he would “not support.”
Back to Claude
His commits to T3 Code show what happened next. Agents add co-author lines to the commits they write, and in July 2026 Claude models start appearing on his commits in large numbers, Fable 5 first, then Opus 5 and Fable 5.1. In October, 463 of his commits name Claude Opus 5.5 and one names an OpenAI model. OpenAI kept a smaller job. In June he posted that asking Opus for a second opinion through claude -p had “significantly improved the quality of the code I get out of OpenAI models.” By October the two had swapped roles: Claude writes, GPT reviews.
Skills
In May he had “almost zero skills installed.” In October he published tstack, “the agent skills I use every day,” for both Claude Code and Codex, and called three of them “the cornerstones of my workflow.” full-send takes a change all the way to a merged pull request: it proves the change, has it reviewed, files the PR and watches it through CI. sol-loop sends the change to OpenAI’s GPT Sol for review until both agents agree, and council asks three models from different providers. The setup needs an authenticated GitHub CLI, the tool he mocked ten months earlier. Days later he made a T3 Code tutorial “on how I use it since I’m still contributing to it,” and in November he walked through how he made Opus “hill climb” until its UI designs looked right.
t3dotgg/tstack2.9kThe agent skills I use every day. Works with Claude Code and Codex.
- skills/
- LICENSE
- README.md
Bring in three delegates to help with the user's request, usually for review
or feedback. Use one strong model from each of three different providers, at
high reasoning effort. If the user's instructions name council members, use
those. Otherwise pick the latest flagship model from each provider, for
example Claude Opus, GPT Sol, and Gemini Pro.
In T3 Code, use `orchestrator_capabilities` to resolve the models, then
`delegate_task` to start one delegate per model in parallel. Elsewhere, use
the harness's subagent tools. If you cannot reach three providers, or a named
member is unavailable, report it instead of substituting.
What the stack says
He still writes TypeScript on a Mac, and his apps use services such as Clerk, where he says he has about 40 projects. The tool between him and the code changes about once a year: autocomplete, then an AI editor, then an app that manages agents, which he now builds himself. The models inside it moved faster, from Claude to GPT and back in about seven months, and because T3 Code runs other companies’ agents, those switches no longer meant a new tool.
Claims I could not verify were left out
Summaries of his October videos describe a private “fleet” repository that syncs his skills and credentials across machines, a proxy that routes every subscription through one endpoint, and a VS Code window he still opens for one repository. Their transcripts could not be read directly, so these are not part of the timeline.