Kanban for AI coding agents: the orchestrators compared

I was working with six terminals on two projects, copying plans and diffs between Claude and Codex by hand, and I said “enough”. I needed a tile manager per workspace, and that’s where Plyloft, the app I build, started. It turns out plenty of other people had the same idea.

An orchestrator for coding agents runs several Claude Code, Codex or other CLI sessions at once, each in its own git worktree, and gives you one place to watch them, read their diffs and merge. They differ on four things that matter more than the feature lists: which agents they run and on whose bill, whether a second model reviews the work, which operating systems they support, and whether there is still a company behind them.

Updated on November 9, 2026: Conductor said on November 4 that it is joining Meta, and Conductor Cloud shuts down on November 20. The parts about Conductor are rewritten, Superset is now called Paradev, Orca and T3 Code are added, and every row of the table was checked again.

Two that shut down, and one winding down

This category has a hard time getting paid. Two examples:

• Vibe Kanban. The company behind it, BloopAI, shut down on April 10, 2026, saying “the vast majority are free users”. The open-source project continued, maintained by the community; the remote services were switched off and recent invoices refunded (shutdown post). Its README now says it is sunsetting (repo).
• Terragon. A cloud service that used your own Claude and ChatGPT subscriptions to open pull requests. It shut down on February 9, 2026 (shutdown page), and its code was released as an open-source snapshot dated January 16, 2026, as is (terragon-oss).

And one is winding down for a different reason. On November 4, 2026, Conductor said it is joining Meta. New users can no longer sign up or download the app, Conductor Cloud shuts down on November 20, and the Mac app gets new models and critical bug fixes until early 2027 (announcement). The dates, and what to do about cloud workspaces, are in its section below.

Agents, plans, review, Windows and price, tool by tool

	Agents	Your own plan	Review by another model	Board	Windows	Price
Conductor	Claude Code, Codex, Cursor, OpenCode	Yes, or API keys	Review action: an agent on the diff	Sidebar of workspaces	No	No new sign-ups since November 4, 2026; winding down
Paradev, formerly Superset	20+ CLIs	Through the CLIs	No built-in second model	Yes	No, planned with no date	Free (1 user); Pro $20/user ($16 yearly); Team $60/user; Ultra $200/user
Orca	Any CLI agent; 34 named	Yes	No; your comments on diff lines go back to the agent	GitHub Projects boards; otherwise a sidebar of worktrees	Yes	Free, open source
T3 Code	Claude Code, Codex, Cursor, Grok Build, OpenCode, Antigravity, fx	Yes, or API keys	No; a model can write a guide to a pull request (beta)	No; a sidebar of threads	Yes	Free, open source
Emdash	35 CLIs	Through the CLIs	No; you compare diffs side by side	?	Yes	Free, open source
Paseo	Claude Code, Codex, Copilot, OpenCode and more	Yes	As a skill, different vendors when possible	No; a sidebar by state	Yes	Free, open source
Nimbalyst	Claude Code, Codex; others in alpha	Yes through the CLI, or API key	An AI review in its pull request mode	Yes	Yes	Free for individuals
Cline Kanban	Claude Code, Codex, Cline CLI, OpenCode	Through the CLIs	No	Yes	Yes, via Node	Free, open source
Claude Squad	Claude Code, Codex, Gemini, Aider, OpenCode, Amp	Through the CLIs	No	No	Sessions fail to start (#275)	Free, open source
Sculptor	Claude Code; the Pi harness	?	No; a review step in its skills	No	No	Free, open source
Vibe Kanban	10+ CLIs	Through the CLIs	No	Yes	Through npx	Free; sunsetting
Plyloft	Claude Code, Codex	Yes, both	Yes, in rounds, findings on the diff	Yes, by what needs you	Windows 11, 64-bit (x64), same as Mac	Free (1 card building at a time); Pro €9; Team €9/seat

What each one does

Conductor

The reference of the category, on the Mac. On November 4, 2026, it said it is joining Meta, and this is the timeline it gave (announcement, changelog):

• November 4, 2026: no new sign-ups. New users can no longer sign up or download the app; existing Pro users keep the Pro features for free.
• November 13, 2026: Conductor Cloud goes read-only. You can still read old chats and export data, but not send new messages. Active subscribers get a refund.
• November 20, 2026: Conductor Cloud shuts down, and Conductor begins deleting cloud user data.
• Until early 2027: the local Mac app keeps working and gets new models and critical bug fixes. Then it starts winding down.

If you have cloud workspaces there, the date that matters is November 20. Conductor gives two ways to keep the work: right-click a cloud workspace and choose Transfer to local, which moves it and its chats to your Mac, or go to Settings → Account → Export data to download all your data. Its team says it will have “something new” to use early next year, and that it can’t share much about it yet.

What the app does, for the people who already have it: each task gets “its own workspace, branch, files, terminal, diff, and review path”, in its own git worktree, for Claude Code, Codex, Cursor and OpenCode (docs). Claude Code runs on your Pro or Max plan, API billing or another provider; Codex on your ChatGPT sign-in or an API key (harnesses). It shows your CI checks and forwards failures to Claude, automerges once CI passes, supports Graphite stacks, and has a Review action where an agent inspects the diff (changelog, review and merge). Conductor Cloud arrived on July 30, 2026 and the iOS app on October 2, 2026. macOS only; its installation page still lists Windows and Linux as not available yet (installation). As of October 22, 2026 it was free locally, with Pro at $50 a month per person and Teams at $60 per user; that pricing page is gone now.

Paradev, formerly Superset

Superset renamed itself Paradev on October 26, 2026, and in the same post announced an $11.5M seed round led by Union Square Ventures (post). The command is still superset. It’s an Electron desktop app that runs more than 20 agent CLIs in worktrees, with diffs and pull request workflows; on Pro it adds automations, remote hosts set up over SSH and an iPhone and iPad app, and on Team cloud workspaces. MIT license since November 4, 2026 (ELv2 before), about 16,100 stars, desktop version 2.0.6 on November 8, 2026 (repo). Its board has the columns Idle, Working, Needs attention, Needs review, Merged and Deleted (workspaces). Review is yours, or an agent you prompt or automate; there’s no built-in review by a second model. Mac, with an experimental Linux build; “Windows is planned with no date” (FAQ). Free for one user; Pro $20 per user a month or $16 billed yearly; Team $60 per user; Ultra $200 per user (pricing).

Orca

Stably AI’s MIT desktop app, at version 1.4.254 on November 8, 2026 (repo). Every task gets its own git worktree, with its own branch, files and agent terminals (worktrees). It works with any CLI agent, in a chat view or as the raw CLI, on the subscriptions you already have; its README names 34. Review is yours: you leave comments on lines of the diff and send them to the agent as one batch (Annotate AI Diff). Its README and docs describe no built-in review by a second model. It commits, pushes and opens the pull request, drafts its title and description, and sends failing checks or review comments back to an agent (commit and push). One agent can split a task and start worker agents in their own worktrees, and agents can run on an SSH host, a remote Orca server or a VM in your own cloud account; Orca “does not sell managed VPS hosting” (ways to run). An iOS and Android companion app is in beta (mobile). Free and open source, on Mac, Windows and Linux (download); the Enterprise page lists no price. It has a guide for migrating from Conductor, which opens Conductor’s worktrees where they are.

T3 Code

An MIT “agent harness control surface”: a server on your machine with an Electron desktop app, a web app and iOS and Android apps, at version 0.1.7 on November 4, 2026, with nightly builds (repo). It drives Claude Code, Codex, Cursor, Grok Build, OpenCode, Antigravity and fx on your own subscriptions, or on an API key if you add a second provider instance. Claude runs through Anthropic’s Agent SDK, and its FAQ says this drains Claude’s 5-hour limit faster than Claude Code itself, a bug it has reported to Anthropic (FAQ). A thread works in the current checkout or in a new worktree, and one prompt sent to several models starts a thread and a worktree for each (threads). It commits, pushes and creates the pull request with a generated title and description, lets you review and merge pull requests, and an agent asked to watch a pull request is woken when a check fails (source control). Review guides, in beta, have Claude or Codex write a tour of a pull request for the person reviewing it. That’s a reading aid: its docs describe no step where a second model reviews the work. Free (“There’s no way to pay us to use T3 Code”); its relay for remote access, T3 Connect, is free for up to 3 machines. On Mac, Windows and Linux.

Emdash

An Apache-2.0 “Agentic Development Environment” from Y Combinator’s Winter 2026 batch, about 6,200 stars, version 1.2.11 on November 6, 2026 (repo). It runs 35 agent CLIs, shows several agents’ changes side by side for you to judge, brings in issues from Linear, Jira, GitHub, Notion, Asana and others, and has SSH, GitHub Actions checks and scheduled automations (docs, providers). Free on Mac, Windows and Linux. Its site now calls this app Emdash Classic and says it “still gets fixes, but new development has moved to Emdash 2.0”, a cloud-first version for teams that has a waitlist and no prices (site).

Paseo

A local daemon with desktop, web, CLI, iOS and Android clients, about 23,600 stars (repo). It runs the official CLIs on your own subscriptions (“Bring your subscriptions”). /paseo-committee has two agents analyze a hard problem independently, preferring “different provider families when possible”, and /paseo-advisor gives a second opinion on a design or a diff without editing files (skills). Both are skills you call, not a step in an interface. Free and open source under Apache-2.0, on Mac, Windows, Linux, iOS and Android (download).

Nimbalyst

The successor of Crystal, deprecated in February 2026 (Crystal). An MIT visual workspace with a kanban of sessions, worktrees, diff review where you accept or reject each change, pull request reviews and visual editors, on Mac, Windows and Linux with an iOS companion (repo). Claude runs either through the Agent SDK, usually billed to an API key, or through the Claude Code CLI on your subscription (#613). Its pull request mode can run an AI review (PR reviews), and a blog post shows Codex reviewing Claude’s work as a manual step (blog). The app is free for individuals; Wiki Teams is $20 per user a month, free during the beta (pricing).

Cline Kanban

A local web app started with npx kanban, a research preview under Apache-2.0, about 1,400 stars (docs, repo). One card per worktree and terminal, with Cline CLI, Claude Code, Codex or OpenCode; comments on lines of the diff go back to the agent; Commit and Open PR; linked cards start when the one before them is done. Free, no account, on any system that runs Node, Windows included.

Claude Squad

A terminal app on tmux with a worktree per session; AGPL-3.0, about 8,750 stars; version 1.0.20 of August 20, 2026 is still its latest release (repo). It runs Claude Code, Codex, Gemini, Aider and other local agents, and can commit and push a branch, but has no board, review, checks or team features. Mac and Linux. There are Windows builds, and the app launches, but creating a session fails on Windows, in an open issue reported on 1.0.17 (#275). Free and open source.

Sculptor

Imbue’s MIT app, an experimental research preview, at version 0.49.0 since October 21, 2026 (repo). Workspaces in worktrees, a pipeline of skills from spec to review, and Create PR, after which a “CI Babysitter” agent fixes failing checks. It integrates Claude Code and the Pi harness (which can sign in with ChatGPT); the Codex CLI isn’t integrated. Free, on Apple Silicon Macs and Linux; no Windows.

Vibe Kanban

A kanban board with worktree workspaces and diffs you can comment on, started with npx vibe-kanban as a local server with a browser interface; about 28,600 stars, Apache-2.0 (repo). It switches between more than 10 agent CLIs. The company shut down on April 10, 2026, and the README now says the project is sunsetting; its latest release, 0.1.44, is from April 24, 2026. An open issue reports runs that stay “running” after their worktree is gone (#1571).

Plyloft

A desktop app for Mac and for Windows 11, 64-bit (x64), with the same features on both. It runs the official claude and codex CLIs in real terminals on your own Claude and ChatGPT plans, and never asks for an API key. Each card gets its own copy of your project, a git worktree.

Before any code, Plan with two AIs has one AI write the plan and the other critique it (Ask about this card & Plan with two AIs). Then one builds; at Finish and check your project’s checks run and the other reviews read-only, in rounds, with findings such as F1 · Serious — … under the lines they name. Try it starts your app from the card’s copy, and Ship it opens a pull request and merges it when its checks pass, or merges on your machine. Review a pull request points the same Reviewer at any open pull request (Review a pull request). The Board sorts cards into Needs you, Not started, Working, On hold and Done, and the Inbox counts what needs you across projects (Inbox & History).

Free builds one card at a time. Pro, €9 a month or €86.40 a year, adds cards in parallel, Build with both and Release several cards; Merge when checks pass is on Free too. Team is €9 per seat a month and adds a shared board on GitHub Issues and shared skills (pricing).

What it doesn’t have: no cloud, no mobile app, no Linux, no issue import from Linear or Jira, and two agents, not twenty.

How to choose

This is my view, from what each project’s own pages say and from the three things I look for myself: the same features on Windows as on the Mac, a second AI reviewing the work, and the agents running on my own subscriptions. I haven’t installed the others, so treat this as a short list to try, not a verdict.

• The most agents for free → Emdash or Orca, or Paradev on a Mac. In all three, the review is yours or an agent you prompt.
• A second opinion from another vendor, on your own subscriptions, with a phone client → Paseo. The review is a skill you call.
• A visual workspace with an iOS companion → Nimbalyst.
• A minimal terminal setup → Claude Squad, on Mac or Linux.
• Claude Code and Codex reviewing each other in rounds from plan to ship, on your own plans, with the same features on Mac and Windows → Plyloft. It’s the one I build, so weigh that.

If you’re on Conductor today, the Mac app keeps getting fixes until early 2027, so you have time to choose. The deadline is for the cloud: transfer or export your cloud workspaces before November 20, 2026. What you’d find the same in Plyloft, and what you wouldn’t, is on coming from Conductor. On Windows, see a Conductor alternative for Windows.

If you’re new to running agents in parallel, start with git worktrees for AI agents: every tool on this page is built on them. For how the two agents Plyloft runs compare, see Codex vs Claude Code.

Plyloft is independent and not affiliated with Anthropic, OpenAI, Meta or any project named here.

If you want Claude Code and Codex on one board, download Plyloft and try it free, one card at a time.