Claude Code UIs compared

Claude Code is a coding agent that lives in a terminal, and a small category of tools has grown up around it: desktop apps, control planes, kanban boards and pipelines that give those agents a surface, isolate them from each other, and help you review what they wrote. This page is a map of that category — the raw terminal-plus-skills workflow, T3 Code, Conductor, and runcastle — and a pointer to the head-to-head pages behind each one.

What this category actually is

None of these tools is a model, and none of them is an agent. They all sit on top of agent CLIs you have already installed and are already paying for — Claude Code first, often Codex, Cursor or OpenCode as well — and they compete on what they put around the agent. That turns out to be a handful of things: a surface you can read a diff in, a way to run more than one agent without them trampling each other's files, and some story for getting the finished work onto a branch you trust.

Two questions separate them more than any feature list. How many agent CLIs does it drive? — one is an opinion, five is a hedge. And how much does it insist on before an agent starts typing? — some tools take your task as given and are deliberately neutral about it; others refuse to start until the task has been argued into a spec. Almost every difference in the table below is downstream of those two answers.

Isolation is the one thing everybody agrees on. Every tool here gives concurrent agents their own branch and their own checkout, usually via git worktrees, because the alternative — several agents editing one working tree — fails immediately and obviously.

The four workflows, side by side

These are the dimensions the head-to-head pages actually established, with sources on each of those pages.

Raw Claude Code + a skills workflow

The baseline, and a genuinely good one. Claude Code skills are SKILL.md files that live in ~/.claude/skills/ or a repo's .claude/skills/; a pack like Matt Pocock's gives you the whole methodology — grill the idea, freeze it into a spec, cut tracer-bullet tickets, implement at the seams, review — for the price of a plugin install. Every rule is a markdown file you can read and rewrite, and it works in any repo on any machine with no server, no container runtime and no port to remember.

What it cannot reach is anything outside one conversation: a skill is instructions inside a session, and parallel features, decisions that must survive next week, and a check that stands between the spec and the merge all live between sessions.

Choose this if… you ship one thing at a time and you are usually at the keyboard while the agent runs; you are still tuning the methodology itself and want to edit a skill and see the change in the session you are already in; you cannot run a container or you are on a locked-down machine; or your work is exploratory — debugging, spikes, reading unfamiliar code — rather than feature-shaped. runcastle vs raw Claude Code + skills →

T3 Code

From Theo and Ping Labs, T3 Code calls itself “the open-source control plane for coding agents”. It drives five agent CLIs — Claude Code, Codex, Cursor, Grok Build, OpenCode — from an Electron desktop app, a web app, and native iOS and Android apps, so you can start and review work from a phone. Threads map to git worktrees, you can switch models mid-thread, and a thread ends with one button that commits, pushes and opens the pull request. It is free, MIT, and it does not resell tokens: your subscriptions stay yours. It is also young and says so — the README's own warning is “We are very very early in this project. Expect bugs”.

Choose this if… multi-model is a hard requirement, or you want to kick off and review work away from your desk, or you juggle several repositories at once, or you simply want less opinion — for a bug fix you can describe in a sentence, a thread plus a diff viewer plus a PR button is exactly the right amount of tool. runcastle vs T3 Code →

Conductor

Conductor is a polished native Mac app for running parallel Claude Code, Codex, Cursor and OpenCode agents “in isolated workspaces on your Mac”. Its unit is the workspace: “Each task gets its own workspace, branch, files, terminal, diff, and review path” — a git worktree underneath, with the app creating the branch, copying the local files you allow, running your setup and run scripts, keeping the diff visible and tracking pull request state until the workspace is archived. That path from task to merged PR is a hard surface to build well, and Conductor is built around getting it right for several sessions at once. The local tier is free with your own subscriptions; cloud workspaces, multiplayer and the team features sit on paid plans. It is macOS-only today: the install page says plainly that it “is not available for Windows or Linux yet”.

Choose this if… you are on a Mac and your work is already shaped — a queue of well-understood bugs, chores and small PRs that does not need to be grilled into a spec — especially if review happens on GitHub with other humans, or you want a team and cloud story rather than a purely local one. runcastle vs Conductor →

runcastle

runcastle is an opinionated programming system layered on Claude Code — the IDE to Claude Code's text editor. Every feature gets a persistent session that walks six phases: ideation, spec, tickets, build, review, shipped. You get grilled on an idea until a spec and a set of dependency-ordered tickets fall out; sandboxed agents burn those tickets AFK in a Docker or Podman container, committing to the feature's branch; you test drive the result on its own port and merge. Gates sit between the phases, block by default, and take an override with a one-line reason that is recorded. It runs entirely on your machine — a Bun server plus a browser UI at http://localhost:4512, MIT-licensed, no account and no hosted backend — on macOS, Windows and Linux, and it requires Claude Code on a paid Claude plan as its engine.

The narrow engine support is the price of the opinion: injected skill packs, phase rules, hooks and its own MCP server are all built on Claude Code's extension points, so unlike the other two apps here there is no model dropdown at all. The pipeline is also real ceremony — on a one-line bug fix it is overhead, and it is meant to be.

Choose this if… the work is a feature rather than a task, the expensive part is deciding what to build, and you want to be absent rather than attentive while it gets built — plus a record of why the code is like that, committed to your own repo. If the terminal is keeping up with you today, you do not need it yet; the honest trigger is noticing that you have become the scheduler. The case against runcastle, in more detail →

Also in the category

These did not get a head-to-head page yet and are listed so the map is not misleading — one sentence each, no deep claims.

• Vibe Kanban — an open-source kanban board over parallel coding agents, each task in its own git worktree, whose own site now says “Vibe Kanban is sunsetting. The project will continue as open source and community maintained.”
• Crystal — an MIT-licensed desktop app for running several Claude Code and Codex sessions in parallel worktrees, deprecated in February 2026 in favour of its successor, Nimbalyst.
• Claude Squad — an open-source terminal app that manages several agents (Claude Code, Codex, Gemini CLI, Aider), each in its own tmux session and git worktree, if you would rather stay in the terminal than open a GUI.

Choosing, in one paragraph

If you want more than one model, or you want to work from a phone, take T3 Code. If you are on a Mac, your tasks are already well-shaped, and your review happens in GitHub pull requests, take Conductor. If you ship one feature at a time with your hands on the keyboard, stay in the terminal with skills and spend nothing. If the expensive part of your work is deciding what to build, and you want tickets burning while you are asleep behind a gate that blocks the merge, that is what runcastle is for. None of that is a commitment: every tool here leaves your repository a normal git repository, so switching costs you an install and nothing else.

Sources and method

Every claim about another tool on this page comes from that tool's own site, docs or README, checked on 10 August 2026; the head-to-head pages carry the full source lists and the exact quotes. This is fast-moving software — Conductor was on release 0.79.0 and T3 Code's own README warned it was very early — so if you are reading this much later, check the originals.

• t3.codes and pingdotgg/t3code — T3 Code positioning, supported harnesses, platforms, licence, project status.
• conductor.build, its docs and pricing — the workspace model, macOS-only availability, plans.
• Claude Code skills docs and Matt Pocock's skills — how the raw workflow is installed and run.
• runcastle's README and CONTEXT.md, plus the docs on this site — every claim about runcastle.

If something here is out of date, or unfair to a tool that is not ours, open an issue and it gets corrected. We would rather lose an argument on this page than win one with a stale fact.