Claude Code agent managers compared
An agent manager runs several coding agent sessions at once, such as Claude Code, Codex or OpenCode, each in its own git worktree, so the agents can change one repo without stepping on each other. They differ in where they run, how you watch the agents, and what happens to the work when an agent is done. Kanzai is a kanban board where the columns do the work: moving a card starts the agent, holds its diff for review, and merges the branch through a queue.
Start using Kanzai for free. No account required.
macOS · Windows x86_64 · Linux AppImage x86_64 · Linux deb x86_64 · arm64 builds
WHY KANZAI
The columns do the work
Drag a card into Development and an agent starts in its own worktree and branch. Drag it into Approved and the merge queue merges it.
Plan from a ticket
Connect Jira or Linear through MCP. The project assistant reads the ticket and your repos, then creates focused cards that point to the relevant code.
One card, several repos
Connect all the repos a project needs. One card can span several of them while Kanzai keeps the branches together.
Review without switching tools
Read the diff, comment on individual lines, and send all your feedback back to the agent in one go.
Merge in the order you choose
You set the order of the merge queue. If something fails, Kanzai stops and shows exactly what made it in.
Make the workflow yours
Add the columns your team uses and set what each one does when a card arrives. Add buttons for the jobs you run all the time.
Sketch it, then write it down
The project assistant draws a diagram you can comment on, and writes the plan into a document. A card can point to the document, and its agent reads it.
Open the code at the right lines
Ask for the method where the user is saved. The file opens at that method, with the lines highlighted.
From ticket to merged code, on one board
Every tool here runs Claude Code agents side by side. Kanzai goes further on the parts around it: planning the work, reviewing it in one place, and merging it in the order you choose.
As of 30 September 2026. Each competitor ✓ and note links its source. A note without a ✓ means partly. A dash means no, or that no public source says so.
| Feature | Kanzai | Conductor | Nimbalyst | Claude Squad | Vibe Kanban | Cline Kanban |
|---|---|---|---|---|---|---|
| Organise the work | ||||||
| Kanban board | ✓ | — | ✓ | — | — | ✓ |
| Your own columns | ✓ | — | ✓Tracker only, not sessions | — | — | — |
| Columns that act when a card arrives | ✓ | — | — | — | — | — |
| Stacked cards: wait, then fork from the card before | ✓ | GitHub PR stacks | — | — | — | Linked, but forks from base |
| One task across several repos | ✓ | ✓ | ✓ | — | ✓ | — |
| Assistant that turns a ticket into tasks | ✓ | ✓Pro only | ✓ | — | — | ✓ |
| Cards grouped by ticket, on its own branch | ✓ | No shared branch | Own tracker only | — | — | — |
| Sketches, docs and files | ||||||
| Diagrams you and the agent share | ✓ | — | ✓ | — | — | — |
| Documents the agent reads | ✓ | Notes folder | ✓ | — | — | — |
| Comment threads with the agent on diagrams and docs | ✓ | — | Teammates only | — | — | — |
| Open any file in the app | ✓ | ✓ | ✓ | — | — | — |
| Comments on files become cards | ✓ | — | Markdown tags only | — | — | — |
| Run the agents | ||||||
| Claude Code | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ |
| Codex | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ |
| Grok Build | ✓ | — | ✓Alpha | — | — | — |
| Antigravity | ✓ | — | Its language server, for Gemini, alpha | — | — | — |
| OpenCode | ✓ | ✓ | ✓Alpha | ✓ | ✓ | Listed, launch turned off |
| Gemini CLI | — | — | — | ✓ | ✓ | Listed, launch turned off |
| Own worktree per task | ✓ | ✓ | ✓Optional | ✓ | ✓ | ✓ |
| Terminal per task | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ |
| Agents keep running when the app is closed | ✓macOS and Linux | — | — | ✓ | While the server runs | While the server runs |
| Take a card over in your own checkout, then hand it back | ✓ | Opens in your IDE | Terminal in the worktree | ✓ | Opens in your IDE | Opens in your editor |
| Review and ship | ||||||
| Line comments sent back to the agent | ✓ | ✓ | — | — | ✓ | ✓ |
| Review rounds that see only the new changes | ✓ | Per turn | Per edit | — | — | Since your last message |
| Merge from the app | ✓ | ✓Via PR only | ✓ | — | ✓ | — |
| Merge queue in the order you set | ✓ | — | — | — | — | — |
| Ask the agent about a failed merge or command | ✓ | ✓ | ✓Git conflicts only | — | ✓Conflicts only | During commit or PR |
| Open pull requests | ✓ | ✓ | — | — | ✓ | ✓ |
| AI review that comments on the diff | ✓ | ✓ | GitHub PRs only | — | In chat | — |
| Review loop that sends its comments back by itself | ✓ | — | — | — | — | — |
| One-click action buttons | ✓ | ✓ | Prompt presets | — | Dev server | ✓ |
| Where you work | ||||||
| Native desktop app | ✓ | ✓ | ✓ | — | — | — |
| macOS | ✓ | ✓ | ✓ | ✓ | ✓ | — |
| Windows | ✓ | — | ✓ | ✓Only via WSL | ✓ | — |
| Linux | ✓ | — | ✓ | ✓ | ✓ | — |
| Any Git host | ✓ | GitHub needed | ✓ | Push to GitHub only | ✓ | ✓ |
| No account needed | ✓ | — | ✓ | ✓ | ✓ | ✓ |
| From your phone | ✓Pro | — | ✓ | — | — | ✓Self-hosted only |
| Price and licence | ||||||
| Price | Free with limits; Pro $19 per year or $3 per month | Free locally; Pro $50/mo | Free; Teams $20/user/mo | Free | Free | Free |
| Open source | — | — | ✓ | ✓ | ✓ | ✓ |
Kanzai puts planning, review and ordered merging on one board, with the Claude or ChatGPT plan, xAI login or Google sign-in you already have.
Sculptor, Claudia (opcode), Crystal and FleetCode are compared in their own sections below.
Conductor
Conductor, from Melty Labs, runs Claude Code, Codex, Cursor and OpenCode sessions in isolated workspaces from one Mac app. The desktop app is macOS only. Each local workspace is a git worktree with its own branch, and paid plans add cloud workspaces that run in a sandbox, which you can also drive over an HTTP API.
It is strong on the path to a merged pull request: inline diff comments go back to the agent, and it opens the PR and follows its GitHub checks until you merge. Cloud workspaces keep running when your laptop is closed, which Kanzai does not offer.
Where Kanzai goes further: it runs on Windows and Linux as well. It works with any Git host, while Conductor needs GitHub. And it puts the work on a board, not in a list of workspaces. Conductor's local use is free, with Pro at $50 a month for the cloud.
Nimbalyst
Nimbalyst calls itself the visual workspace for Codex and Claude Code. It is a desktop app for macOS, Windows and Linux with an iOS companion, and it runs Codex and Claude Code side by side, with OpenCode, Copilot, Grok Build and Gemini as further providers in alpha. Sessions can be isolated in git worktrees, which you can review, commit and merge back from the app.
It does a lot well. It is MIT licensed, local first with an optional account, and has a session board where you drag sessions between columns. Its visual editors for markdown, mockups and diagrams go well beyond what Kanzai does.
Where Kanzai goes further: every card gets its own worktree and branch. In Nimbalyst, worktrees are optional. Kanzai's columns also do the work, not just track it. Development starts the agent. Review holds the diff. Approved merges it through a merge queue. Nimbalyst is free for individuals, and Teams costs $20 per user per month.
Crystal, now Nimbalyst
Crystal has been deprecated and replaced by Nimbalyst, from the same makers. Its last release, in February 2026, recommends migrating, and says Crystal still works but will not be updated.
It was a desktop app that ran Claude Code, Codex or both, each session in its own git worktree, with a diff viewer, a commit per iteration, and a rebase-to-main step. It is MIT licensed. Nimbalyst describes it as the first GUI for managing multiple Claude Code sessions in parallel.
If you used Crystal, its makers point to Nimbalyst, above. Kanzai is the option if you want each task on a board card, with review and merge as columns.
Claude Squad
Claude Squad is a terminal app that manages multiple Claude Code, Codex, Gemini and Aider agents, each in a tmux session and its own git worktree. Release builds cover macOS, Linux and Windows, where tmux comes from WSL or similar.
It is light: one terminal window, any agent CLI as a profile, tasks that run in the background, including auto-accept mode, and a diff tab to review before you commit and push. It is free and open source under AGPL-3.0.
Where Kanzai goes further: it is a desktop board with a transcript, a diff you comment on line by line, and a merge queue. If you want to stay in tmux, Claude Squad is the closer fit.
Vibe Kanban
Vibe Kanban starts with npx and opens its UI in your browser. Each workspace gets git worktrees for the chosen repos, and it supports ten coding agents, from Claude Code, Codex and OpenCode to Gemini CLI, Copilot and Amp. It runs on macOS, Linux and Windows.
On 10 April 2026, Bloop, the company behind it, shut down. The project continues as open source and community maintained, and a community release came out in September. The same change replaced the interactive kanban board with an export screen, though local workspaces still work.
Its strengths are breadth: inline review comments sent back to the agent, pull requests through the GitHub CLI, and a built-in browser preview with devtools, all open source under Apache-2.0. Kanzai is a native app rather than a local web server, runs five agents to its ten, and keeps its board as the way you drive the work. There is more on that on the Vibe Kanban alternative page.
Cline Kanban
Cline Kanban, from the makers of Cline, starts a local server and opens the board in your browser, with no account. Each task card gets its own terminal and worktree, and it works with Cline CLI, Claude Code, Codex and other runtimes, with a different agent per task if you like. Cline calls it a research preview.
It is good at running work unattended: a linked card starts on its own when the one before it completes, and diffs are scoped to message checkpoints, with line comments sent back to the agent. It is open source under Apache-2.0.
Where Kanzai goes further: it is a native app. You can add columns of your own; Cline Kanban has four fixed ones. Kanzai merges approved branches itself, in the order you set. Cline Kanban asks the agent to commit or open a pull request.
Sculptor
Sculptor, from Imbue, is a desktop app for running coding agents in parallel, with builds for Apple Silicon Macs and Linux. A workspace is a git worktree by default, and a Docker backend is experimental. It supports Claude Code and the Pi harness, and can run any terminal agent.
It opens a GitHub pull request and tracks its reviews and CI, keeps an interrupted Claude turn across a crash, and ships workflow skills such as a fix-bug flow that writes failing tests first. It is MIT licensed, and Imbue calls it a research preview.
Where Kanzai goes further: it runs on Windows, Intel Macs and Linux. Sculptor has no Windows or Intel Mac build today. Kanzai also runs Codex, Grok Build, Antigravity and OpenCode, and puts every task on a board.
Claudia (opcode)
Claudia was renamed opcode in version 0.2.0. It is a GUI app and toolkit for Claude Code, built with Tauri. Prebuilt downloads cover macOS and Linux, and Windows builds from source.
It focuses on the Claude Code session itself: checkpoints you can fork and restore, a diff between checkpoints, custom agents, and a usage and cost dashboard. It is free under AGPL-3.0. The latest release is from August 2025.
Where Kanzai goes further: every task gets its own worktree and branch, so several agents can change one repo at once. opcode runs agents as separate processes and does not mention worktrees. Kanzai also runs Codex, Grok Build, Antigravity and OpenCode.
FleetCode
FleetCode is a desktop terminal application for running several CLI coding agents at once, each in its own git worktree. It runs Claude Code or Codex, and its beta builds cover Apple Silicon Macs and Linux.
It stays small: the real CLI in a terminal per worktree, sessions that resume after a restart, per-session setup commands, and MCP server settings. It is free under the ISC licence.
Where Kanzai goes further: finished work waits in Review for your line comments. Approved branches merge through a queue. FleetCode's README describes no review or merge step.
- Native app for macOS, Windows and Linux
- Runs locally, with no account and your data in one file
- Runs the real Claude Code, Codex, Grok Build, Antigravity and OpenCode CLIs
- Use your Claude or ChatGPT plan, xAI login or Google sign-in; only OpenCode may need an API key
- Works with any Git host
Try Kanzai on your own repo.
Run Claude Code, Codex, Grok Build, Antigravity and OpenCode side by side from one local board. Drag cards to start agents, reviews, and merges. Keep using the subscription you already pay for.
The button picks x86_64 for Windows and Linux. On arm64 hardware, take the arm64 build.
Kanzai is an independent product and is not affiliated with any of the tools named here. Names are trademarks of their owners.