Open-source Claude Cowork alternatives, compared for a team

Yes: four open-source projects do the work Claude Cowork does, and three of them shipped a release this month. The listicles ranking them are written by the alternatives themselves. eigent.ai's guide puts Eigent first, kuse.ai's opens with Kuse, and a directory page still dated 2025 rounds out the first page.

A list written by a competitor answers the questions its author wins at. For a team the questions come first: where the company's way of working lives, who reviews what an agent did before it counts, where the work runs, and what you keep. Kortix, the open-source option here, holds the company's setup in a git repo the whole team owns. The four alternatives keep that setup close to one person's machine. For a team, our recommendation is Kortix.

The field at a glance

All five are open source, which already separates them from Cowork itself. Where they split is where the work runs and where the setup lives.

The first column is easy to check. The last two decide what the team keeps once the person who set everything up moves on.

Two names stay out of the table on purpose. OpenAI's version of the category is ChatGPT Work, shipped in July 2026: closed, usage-metered, on OpenAI's cloud. No open-source build sits behind it (Kortix on GitHub keeps the running comparison). Cowork itself is no longer a separate product either. Anthropic folded it into Claude on September 16, 2026 (release notes). The merged experience is reaching Pro and Max plans first, and there is no Cowork-only free tier to sign up for. Pro runs $17 a month billed annually or $20 month to month, checked October 2026 (Claude's pricing page).

Kortix: the system a team runs

Kortix, the open-source AI Operating System, runs the company from one git repo the team owns: agents and their skills, the memory they share, and every connector they use. Open source. Read it, fork it, audit it.

Every project is a git repository, with kortix.yaml at the root of it. Agents, skills, memory, tools, triggers and connectors all live in it as files (Kortix docs). One session runs one agent, alone in a sandbox of its own, on its own branch of the repo. To keep any of its work, the agent commits and asks for a merge through a change request, which a person with the merge grant reads and merges. Agents hold no merge rights until the project grants them.

Run it on Kortix Cloud, in your VPC, or on your own on-prem network. Self-host is free. Models: any model provider with your own keys. Or the ChatGPT plan you already pay for. Free is $0 with 200 credits per month and 1 project. Paid plans start at $50 a month for 5,000 credits, and members are free, sharing one credit pool (Kortix pricing).

The team runs its own company on it. Counted from the repo's public git history over the last 30 days, Kortix Agent is the most frequent author on the open-source repo, ahead of every person (Kortix on GitHub).

The honest cost is setup. Self-hosting runs the full stack on your infrastructure, with a sandbox provider behind the sessions, and that is more work than a desktop installer (self-hosting).

OpenWork: the polished desktop with an enterprise lane

OpenWork describes itself as "a desktop app for macOS, Windows, and Linux where AI agents do real work on your own files", built on OpenCode (its repository). The desktop app is free and MIT-licensed. Models are open: 50+ providers on your own API keys, or local models through Ollama. For organizations there is OpenWork Den, a control plane you can self-host, which publishes skills and plugins through marketplaces, manages access, and sets desktop policies. The license is directory-split: everything outside ee/ is MIT. The control plane under ee/ sits under the OpenWork EE License, and production use of it requires a subscription outside the exemptions, which include organizations of up to five people. On OpenWork Cloud, the Team plan includes the first five seats free, then $10 per seat per month, checked October 2026 (openworklabs.com).

The team gap is the desktop itself. The agent works on the files of the machine the app runs on, and neither the site nor the repository describes a review gate between an agent and anything shared.

Open Claude Cowork: the quickest start, quiet since May

Composio's project is "an open-source desktop chat application powered by Claude Agent SDK and Composio Tool Router": an Electron app with 500+ SaaS integrations, under the MIT license (its repository). You bring an Anthropic API key and a Composio key, and model choice runs through the Claude Agent SDK or Opencode. Setup is a script and two terminals, which makes it the quickest thing in this field to clone and run.

It is also the quietest. The default branch last changed in May 2026, with no releases, checked October 2026. For one person that is a fair trade for a fast start. For a team, an unmaintained repo is a risk of its own.

Eigent: a multi-agent desktop, team features still landing

Eigent's repository calls it "the open source Cowork desktop application": a multi-agent workforce built on CAMEL-AI's framework, under Apache 2.0 (its repository). A local standalone mode needs no cloud account, and models connect by your own API keys or run locally. Their site makes the team pitch: deploy locally or self-host in an environment your team controls, manage roles and access, and approve the result or ask for revisions (eigent.ai). Two of their own pages mark the distance to those promises. The pricing page lists shared library, centralized billing and seat management as Coming Soon. It says SOC 2 readiness and GDPR compliance work remain in progress. The Pro plan runs $19.99 a month after a 7-day trial, for 2,000 task credits, checked October 2026.

Where it stops for a team: the coordination happens on one machine. Memory is scoped, in their words, to you, the Space, or the Session.

OpenClaw: the assistant that lives in your chat apps

OpenClaw is an open-source assistant built to run on your machine and meet you in the channels you already use, from WhatsApp and Telegram to Slack and Teams (openclaw.ai, the project's repository). It is MIT-licensed, stewarded by an independent 501(c)(3), with no subscription and no hosted tier. The ownership promise is the site's own list: "Your memories. Your skills. Your models. Your machines. Your data." Its repository adds that one gateway can run it as a personal assistant on a laptop or as a shared team deployment. The configuration is the only difference.

The shape is the limit here. The product is an assistant for a person, and its pages describe no company repository, no review gate, and no isolated machine per task.

The axis the lists skip

A desktop app is the right shape for one person's files, and all four of these projects build that well. A company is a different shape. Its work spans people and machines, its way of working has to outlive any one laptop, and someone has to be able to answer for what an agent did.

Kortix answers them with files, and the files are public. Two files in every starter project make the case. The manifest, kortix.yaml, opens:

# Kortix project manifest. Source of truth for project-wide config that
# lives in git alongside the code.

The same file grants nothing by default. The agents section reads: v2 is deny-by-default: an omitted connectors/secrets/kortix_permissions/skills grant resolves to "none", so a scoped specialist starts locked down. And agents/kortix.md, the file a new session's agent reads, defines what a session is:

The session you're in is an isolated sandbox — an ephemeral branch of
this repo, your own `/workspace` — so you can install, experiment,
and recover freely. Only what you commit + push survives.

The company's rules sit next to its code, versioned and diffable. A person can read them, and an agent can propose changes to them. A session is a branch that vanishes unless its work is committed, and nothing lands on main without a diff a person approves.

The four alternatives answer the same questions in their own words. OpenWork: your files stay on your machine. Eigent: secrets stay bound locally. OpenClaw: your memories, your skills, your machines. Open Claude Cowork keeps sessions in a local Express server. Skills can be shared across a team in some of them, but none of them holds how the team works in one versioned repo.

When to pick which

For a team, pick the system the team can own: Kortix. The company's agent definitions, its skills, its memory and its connector rules sit in one repo, and a person approves every change before it lands. The work runs on cloud computers instead of one laptop, and any model works with your own keys. Kortix is also the heaviest of the five to stand up, the price of owning the whole system.

The desktop cases are real. If you are one person who wants Cowork's desktop feel without Anthropic's terms, OpenWork is the pick. If you want an always-on assistant that lives in WhatsApp or Slack, OpenClaw. If you want a multi-agent workforce on one machine, Eigent.

The head-to-head with Cowork itself, Kortix vs Claude Cowork, covers what changed when Anthropic folded it into Claude. One layer down, five open-source platforms compared for a team sorts frameworks from application platforms. One kortix init starts a project. Get started with Kortix.