skribbl
productpricingfree!questionswriting
download
comparisons · 6 August 2026 · 8 min read

Open-source agent orchestration tools, and their gaps

Every one whose licence we could verify, what it is good at, and where the facts run out.

skribbl/writing/comparisons
what each one is actually for

The open-source tools for running several coding agents at once: what each one does, how it isolates agents from each other, what it costs to run, and the kind of work it suits. Read off each project repository and release feed, with the date of the reading stated and the gaps left as gaps.

the short version

Of the parallel-agent tools we track, four state an open-source licence on their own repository: Vibe Kanban, Crystal, Nimbalyst, Claude Squad. One more, Sculptor, belongs in the same conversation but publishes no licence we could confirm, so it is described below and not counted. All of them run several coding agents against one repository at the same time, and they differ mainly in interface shape and in how hard a wall they put between the agents.

Nothing here was benchmarked. Every fact below was read off each project's own repository, documentation and release feed on 2 August 2026, and where a project does not document something, this post says the documentation is silent rather than claiming the feature is missing. Two of them are no longer where their authors are working, which is said plainly in their sections rather than left for you to discover.

VIBE KANBANa board you groom and dispatch from, driving more agent CLIs than anything else here
CRYSTALparallel sessions built around trying one task several ways and diffing the results
NIMBALYSTthe successor project: a visual workspace holding documents and mockups beside a board of sessions
CLAUDE SQUADthe same job done entirely inside a terminal, on any machine that Go runs on

vibe kanban

A kanban board for planning work and dispatching it to coding agents, each task running in its own isolated workspace with a branch, a terminal and a dev server.

Status: unmaintained. Bloop shut down 10 April 2026. Last release 24 April 2026. The source is open, so the code does not go anywhere, but nobody is shipping fixes to it today and you should read the rest of this section with that in front of you.
SHAPEKanban board
PLATFORMSAnywhere Node runs, via npx. Browser UI
PRICEFree
SOURCEOpen, Apache-2.0
AGENTS DRIVENTen or more, including Claude Code, Codex, Gemini, Copilot, Amp, Cursor
SPEND METERNot documented
WORKTREESYes, with cleanup of orphaned ones
STATUSunmaintained. Bloop shut down 10 April 2026. Last release 24 April 2026

what it is good at

These are the concrete advantages we recorded for it, each one a thing it does better than the tool this site sells:

  • Apache-2.0 and self-hostable. It runs on any operating system, there is no vendor to lose, and it can be forked forever. Skribbl is a closed macOS binary with a licence key.
  • A backlog you groom and dispatch from is a genuinely different and better model for queued work than a canvas of live terminals. If your work arrives as tickets, this shape fits it and ours does not.
  • It drives more agent CLIs than anything else here, and pairs inline diff review with a built-in browser preview that has device emulation.

the work it suits

Work that arrives as tickets. If your day already starts with a backlog and ends with things moved to done, a board is the shape your work is in, and dispatching a card to an agent is one motion rather than a decision about which terminal to type in. It runs wherever Node runs and the interface is a browser tab, so the machine doing the work does not have to be the machine you are sitting at.

Anything where you need somebody to fix a regression. The release feed stopped and the company behind it is gone, so adopting it is adopting a fork: the licence gives you the right to maintain it and nobody else is going to. That is a fair trade for a stable internal tool and a bad one for something you want to still work after the next major CLI release.

crystal

A desktop app for running several Claude Code and Codex sessions at once in parallel git worktrees, to try an approach more than one way and compare the results.

Status: superseded. Discontinued, replaced by Nimbalyst. Last release 26 February 2026. The source is open, so the code does not go anywhere, but nobody is shipping fixes to it today and you should read the rest of this section with that in front of you.
SHAPESession list
PLATFORMSmacOS universal, including Intel. Linux
PRICEFree
SOURCEOpen, MIT
AGENTS DRIVENClaude Code, Codex
SPEND METERCould not verify
WORKTREESYes, central to it
STATUSsuperseded. Discontinued, replaced by Nimbalyst. Last release 26 February 2026

what it is good at

These are the concrete advantages we recorded for it, each one a thing it does better than the tool this site sells:

  • MIT, free, and readable. You can check what it does to your repository rather than take a stranger’s word for it.
  • It runs on Intel Macs and on Linux. We ship one build, for Apple Silicon, and that rules out a real number of people.
  • Running the same task several ways and diffing the outcomes is a workflow it serves directly. Skribbl is built for a fleet doing different work, not the same work in triplicate.

the work it suits

The case where you do not know the approach yet. Three sessions attempt the same awkward migration, you read three diffs, and you keep the one that is least bad. It is also the answer for hardware the rest of this list ignores, because it ships for Intel Macs and for Linux.

New work you expect to still be building on in a year. It is readable and it still runs, but the project has handed over, and starting there today means starting on a codebase whose author has moved to the next one.

nimbalyst

Crystal’s successor: a visual workspace holding markdown, diagrams, mockups and a browser preview beside a kanban of agent sessions.

SHAPEVisual workspace and session board
PLATFORMSmacOS including Intel, Windows, Linux
PRICEFree
SOURCEOpen, MIT
AGENTS DRIVENClaude Code, Codex
SPEND METERNot documented
WORKTREESYes, optional per session
STATUSactive. Shipping actively

what it is good at

These are the concrete advantages we recorded for it, each one a thing it does better than the tool this site sells:

  • It is free, MIT, and on all three desktop platforms. Of everything on this page it is the one closest to what we are building, and it costs nothing.
  • A visual workspace that holds diagrams, mockups and documents alongside sessions. Skribbl’s canvas holds terminals and files, not artefacts you author.
  • Mobile companions. We have none and none planned.

the work it suits

Work where the thinking and the running belong in the same place. Notes, a diagram, a mockup and the sessions acting on them sit on one surface, which fits design-led work and anything where the plan changes while the agents are mid-task. It is the only tool in this group covering all three desktop platforms.

A workflow built on a CLI it does not drive. The agent list is short, so if your team standardised on something outside it, that is the constraint to check first. Nothing in the documentation describes running cost, so budget stays a question you answer from your provider dashboard.

claude squad

A terminal program that manages several agents at once in isolated worktrees, without leaving the terminal.

SHAPETerminal interface
PLATFORMSAnywhere Go runs
PRICEFree
SOURCEOpen, AGPL-3.0
AGENTS DRIVENClaude Code, Codex, OpenCode, Amp, Aider
SPEND METERNot documented
WORKTREESYes
STATUSactive. Shipping actively

what it is good at

These are the concrete advantages we recorded for it, each one a thing it does better than the tool this site sells:

  • It never asks you to leave the terminal, which for a lot of people is the entire requirement and the reason a desktop app is a downgrade.
  • It is free, AGPL-3.0, and runs over SSH on a machine with no display at all.

the work it suits

Remote and headless work. It runs over SSH on a box with no display, which is exactly the setup people reach for when the repository is large, the tests are slow, or the laptop is not where the work should happen. For anyone who already lives in tmux, a desktop app is a step sideways and this is not.

Watching a lot of things at once. A terminal interface shows you what you have scrolled to, and past a handful of agents the question "which one is waiting on me" gets harder rather than easier. Its documentation does not describe a spend meter either, so that is a gap you fill yourself.

sculptor, and why it is not on the list

A desktop app that runs parallel Claude Code agents in Docker containers, each with a full copy of the repository, and syncs any agent’s work into your editor live.

It is the obvious fifth entry and it is not counted as one, for a boring reason: its source field reads could not verify. We could not find a licence published on a primary source, and a roundup that guesses at licensing is worse than a roundup with a gap in it. If that changes, this section changes. Everything else about it is documented, so here it is on the same terms as the four above.

SHAPESession list
PLATFORMSmacOS Apple Silicon, Linux
PRICEFree while in beta
SOURCECould not verify
AGENTS DRIVENClaude Code
SPEND METERNot documented
WORKTREESNo. Containers, with a full repository copy each
STATUSactive. In beta, built by Imbue

what it is good at

  • Containers are stronger isolation than worktrees. An agent that runs a destructive command hits a container rather than your machine, which is a real answer to a real fear that we do not have.
  • Its pairing mode pulls whichever agent you are watching into your own editor live.
  • Free while it is in beta.

the work it suits

Work you are nervous about letting an agent do unattended. It is the only tool here that puts a real boundary rather than a directory boundary around each agent, which changes what a badly chosen shell command can reach. It also drives a single agent CLI, so it fits teams that have standardised on one and not teams mixing several.

how each one isolates agents from each other

This is the axis that actually separates these projects, and it is worth more attention than the interface. Two agents in one working directory are two processes writing the same files with no lock between them, and the failure is quiet: one reads a file, thinks for a minute, and writes back a version that discards what the other did in between.

There are three answers in this group, and every project here has picked one.

VIBE KANBANYes, with cleanup of orphaned ones
CRYSTALYes, central to it
NIMBALYSTYes, optional per session
CLAUDE SQUADYes
SCULPTORNo. Containers, with a full repository copy each

A git worktree per agent is what most of them do. It is cheap, because several checkouts share one object database, and it puts real conflicts where they belong: at merge time, in front of a person. It does nothing about a stray rm or a migration run against the wrong database, because a worktree is a directory and not a boundary.

A container per agent is the stronger version. A full copy of the repository, a process that cannot reach your machine, and a real answer to the thing people are actually afraid of. It costs disk and it costs setup, and syncing work back out of a container is a problem worktrees never have.

Nothing is the third answer, and it is the default if you run several agent CLIs in several terminal tabs yourself. It works right up until it does not. If you would rather do the isolation by hand and keep the tooling out of it, the commands are short and we wrote them out step by step in how to set up a multi-agent coding workspace. Doing it manually is a legitimate choice; doing it by accident is not.

what free actually costs

Every project above is free to install, and none of them is free to run. Open source removes the licence fee. It removes nothing at all from the model bill, because none of these tools ships a model: they drive Claude Code, Codex and the rest, and every token those spend is billed to your account by whoever provides them.

The part that catches people is that the burn rate is per agent. One agent working unattended for twenty minutes costs what it costs. Four agents working unattended for twenty minutes cost roughly four times that, in the same twenty minutes, and nothing in a terminal tells you while it is happening. That is the real difference between a free tool and a cheap afternoon.

Estimating it from token counts is harder than it looks, too. In a long coding session the overwhelming majority of tokens are cache reads rather than fresh input, and cache writes are priced differently again, so summing raw counts against one rate gives a number that is wrong by a multiple rather than by a margin. We took that apart with the arithmetic in what Claude Code actually costs.

Check what your chosen tool documents about spend before you run a fleet on it, not after. In the data above, the meter row for most of these projects reads "not documented", which means the project does not describe one, not that it lacks one. Either way, plan to get your running total from your provider rather than from the tool.

how this was checked

Every fact in this post was read off the project's own repository, documentation or latest release on 2 August 2026. Nothing was quoted from another roundup, and the date is printed from the same file the facts live in, so it cannot drift away from them.

We did not benchmark anything. There are no timings, no star counts, no install sizes and no scores in this post, because we did not run these four side by side under matched conditions and any number we published would be decoration. What you get instead is what each project says about itself, checked, with the gaps left visible.

Where a project does not document a feature, the entry says so in those words. "No spend meter documented" and "has no spend meter" are different claims and only the first one is ours to make. Where we could not confirm something at all, the entry says it could not be verified, which is why one obvious candidate is described above and not counted among the open-source projects.

Licences, prices, platforms and maintenance status all move, and this group moves in weeks. Between two drafts of the comparison table this post reads from, one project shut down and another handed over to a successor. Treat 2 August 2026 as an expiry date and check the repository before you commit an afternoon.

If you maintain one of these and something here is wrong, tell us and it gets corrected in the data, which corrects this page and the comparison table together.

our own position, stated plainly

Disclosure. Skribbl, the product this site sells, is a commercial tool in the same category as the projects above. It is not open source: it is a closed macOS binary with a paid licence, and we are not going to pretend that a post recommending free alternatives is disinterested. It is not on the list because it does not qualify, and if the only thing that matters to you is a licence you can read and fork, one of the four above is the right answer and we are not a candidate at any price.

The full table, including our own row with the parts that lose, is at /compare. It is built from the same data file this post reads, so the two cannot disagree with each other.

READ NEXT
What is an AI agent orchestration platform? 2026 guideTwo families, seven capabilities to check, and when to buy nothing at all.10 minLangChain vs CrewAI vs AutoGen, compared honestlyThree mental models, honestly compared, plus the question none of them answers.10 minClaude Code GUI alternatives, and why people switchWhat breaks at four agents, and the four shapes of GUI people reach for.9 min
ON THIS PAGE
the short versionvibe kanbancrystalnimbalystclaude squadsculptor, unconfirmedhow each one isolates agentswhat free costshow this was checkedour own position
run them on a canvasSkribbl puts every agent, its terminal and what it is spending on one board. macOS, one day free.

get the next one by email.

One email when there is something worth reading. Unsubscribe is one click and it is in every issue.

get me
productpricingdocsquestionswhat it iscomparereleaseswritingnewsletterlaunchesprivacycancel
give them infinity.© skribbl