Compare
Trek vs T3 Code
Both let you bring your own coding agents and run them on a machine you connect to. They're built for different journeys. T3 Code takes a single thread from prompt to merged pull request. Trek coordinates a whole expedition of real agent CLIs across branches and repositories, local models included.
Checked against T3 Code's README and user docs in November 2026. Spot something out of date? Open an issue.
Choose Trek if you want
- The real
claude,codex,geminiandgrokCLIs, with every slash command, skill, plugin and hook - Local models through Ollama or LM Studio
- One agent leading the others from a built-in panel
- Projects that span several repositories
- Workflow tags, Linear tickets and notes on every branch
- A server you run yourself, with no account, relay or telemetry
Choose T3 Code if you want
- A guided path from prompt to merged PR, with diffs and generated commit messages and PR descriptions
- PR watching that wakes the agent on failed checks, reviews or conflicts
- GitLab, Bitbucket, Gitea, Forgejo or Azure DevOps
- iOS, Android and web clients today
- A wider list of agents, including Cursor, OpenCode and Pi
- A desktop app on Windows and Linux, installed with brew, winget, AUR or Nix
Side by side
| Trek | T3 Code | |
|---|---|---|
| Agents | ||
| How agents run | The real CLI in a terminal (PTY) | Claude through Anthropic's Agent SDK |
| Claude's 5-hour limit | Used as it is by Claude Code in your terminal | Its FAQ: Agent SDK use "will drain your 5h limits faster than using Claude Code directly" source |
| Agents | Claude Code, Codex, Gemini CLI, Grok, local models | Claude, Codex, Cursor, Grok Build, OpenCode, Pi, Antigravity and more source |
| Local models | Ollama or LM Studio, through Claude Code | Not supported, per its FAQ source |
| Coordination | ||
| One agent driving others | Built-in Global panel with trek-dev MCP tools: create worktrees, start sessions, send and queue prompts, read terminals, wait, ask you to decide | Outside agents can start, message, wait on and stop threads through its CLI and MCP source |
| Several repositories in one project | Yes | Not documented |
| Dev stack per worktree | Yes, on your Mac | Project scripts with a preview URL |
| Tracking work | ||
| Workflow tags on branches | In Progress, In Review, Ready to Deploy and more, with a sidebar filter | Not documented |
| Issue tracker | Linear issue status and title on each branch | Not documented |
| Branch notes | Kept with the branch, even after its worktree is deleted | Not documented |
| Git and pull requests | ||
| Diffs | A Changes view of each worktree | Diff pane per turn, branch or commit |
| Generated commit messages and PR descriptions | Not yet | Yes source |
| PR watching | Not yet | Checks every two minutes; wakes the agent on failed checks, reviews and conflicts |
| Stacked PRs and line-comment reviews | Not yet | Yes |
| Git hosts | GitHub (check out a PR as a worktree, link PRs to branches) | GitHub, GitLab, Gitea, Forgejo, Bitbucket, Azure DevOps and more |
| Worktree cleanup | Manual | Manual or automatic |
| Remote, platforms and privacy | ||
| Server | trek-server on macOS, Windows or Linux | macOS, Windows or Linux |
| Remote access | Straight to your server over Tailscale, your LAN or wss://, one token per device |
T3 Connect over a Cloudflare tunnel ("free too, for now", machine limits by plan), or LAN and Tailscale pairing source |
| Telemetry | None | On unless you set T3CODE_TELEMETRY_ENABLED=false source |
| Mobile and web | iOS and Android coming soon | iOS, Android and app.t3.codes |
| Desktop app | macOS (Apple Silicon) | macOS, Windows and Linux; brew, winget, AUR, Nix, npx t3 |
Yes Partly, or with a catch No Not in its docs
Real CLIs
The real guide, in a real terminal.
Trek starts claude, codex, gemini and grok
in a real terminal, the same way you would. Slash commands, skills, plugins, hooks and
permission prompts behave natively, and Claude Code uses your subscription exactly as it
does in your own terminal. T3 Code drives Claude through the Agent SDK, which its own
FAQ says drains the 5-hour limit faster.
Local models
Bring a guide who works offline.
Some trails are private, or just don't need a frontier model. Trek runs Claude Code against Ollama or LM Studio, so a local model can take a session like any other guide. T3 Code's FAQ says it doesn't support local models.
Expeditions
A whole expedition, not one thread.
A Trek project can hold several repositories. Every branch carries a workflow tag, its Linear issue and notes that outlive the worktree, and the Global panel's agent can spread work across worktrees, queue each session's next leg and wait for them all. T3 Code is built around threads that each run from prompt to pull request.
Camp
A camp that answers to you.
trek-server runs on your own machine and talks straight to your devices: no account, no relay and no telemetry. T3 Code offers LAN and Tailscale pairing too, and its easiest remote path, T3 Connect, runs through a Cloudflare tunnel with machine limits tied to your T3 Chat plan.
Where T3 Code is ahead today
T3 Code is an excellent, popular open-source (MIT) project, and on some ground it's further along. Its pull request workflow is deeper, with per-turn diffs, generated commit messages and PR descriptions, stacked PRs and review with line comments, and its server watches PRs and wakes the agent when checks fail or reviews arrive. It works with many more git hosts and agents, its iOS, Android and web clients are out now, and it installs through every major package manager. If that's the route you need, it's a great choice.
Where will yours lead?
Run your real agents across every branch and repository, with a camp of your own.