Claude Cowork vs Eigent: open source on a laptop or a repo

Eigent, and the machine it runs on

Eigent is an open-source desktop app that runs a team of agents on the computer where you install it. It is built on CAMEL, an open-source multi-agent framework, with an Electron front end and a Python backend, and its documentation calls it the first open-source general-agent desktop app (Eigent docs). Work is arranged as Space, Session, Task: a Space can point at a local folder, a Session holds related work, and a Task is one request handled by a single agent or by a workforce of agents running in parallel. The app ships more than 200 MCP tools, a skills system, scheduled and event automations, and it asks a person for input when a task gets stuck.

The requirements are on its self-hosting page: a desktop system or a Linux environment, git, Node 18 through 22, the Python version the backend needs, disk for the dependencies and any local model, and either a provider key or a running local model server (self-hosting). Those are the terms of a desktop product. In exchange the files, the credentials and the inference can all stay on hardware you control, and the model can be Ollama, vLLM, LM Studio, a hosted API or an enterprise gateway.

Where that leaves a team is the part worth saying plainly. The app is the runtime: a run happens inside it, on the machine it is installed on, so the machine has to be awake for the work to finish. The Spaces, connector accounts and keys that make those agents useful belong to that install, and the colleague who set it up is the one who knows where each piece is. Eigent also offers a cloud platform that hosts the models, APIs and storage while the app stays on the desk, and it lists Eigent Cloud as the arrangement for model execution, with Enterprise adding local deployment, SSO and access control (Eigent). None of that turns the install into a shared, reviewable record of what the company runs.

What Claude Cowork became in September

On 16 September 2026 Anthropic merged Claude Cowork and chat into one Claude, so what Cowork does is now reachable from any conversation, with the context, skills and connectors already in the account. It rolled out to Pro, Max and Ultra plans first, with more plans to follow (the announcement).

The behavior is documented on the product page. Cowork works in folders and tools you choose, opens its own browser in a side panel instead of taking over yours, keeps running after you close the laptop, and splits a large project into chunks that run together. New tasks start in "Automatically approve" mode on Pro, Max and Team plans, and in "Manually approve" on Enterprise plans. An organization owner can turn the mode off for everyone, and an admin can set access by team, spend limits and tool permissions by department, with activity streaming to a SIEM through OpenTelemetry. It runs in Anthropic's cloud or in a customer's Amazon Bedrock, Gemini Enterprise Agent Platform or Microsoft Foundry account, and Anthropic documents no way to host the product yourself (Claude Cowork).

What a seat costs is public too: Pro at $17 a month, Max at $100, Ultra at $200, Team at $20 a seat, Enterprise quoted, with usage limits applied to each.

The setup is the thing that leaves

Both hand back documents, both can schedule a task, and both put a person in the loop before the work lands, so the quality of the work is a wash. What separates them is where the work is configured: an account with a plan attached, or an install on a laptop.

Run a test on the last job your team automated. Who can say which instructions it follows, what it may send and delete, and what it learned last quarter? If the honest answer is a colleague's laptop or a colleague's account, the company does not have that job. It has a person who has that job. The same gap shows up at review time: an approval inside one app is a click that leaves no artifact, while a diff in a repository is a record a second person can read six months later, after the author has moved to other work.

That is the axis the two products share, and neither of them crosses it. A desktop app and a subscription seat are both ways of running agents well. A company needs the third thing as well: somewhere the agents, their skills and their memory live as files that survive the person, the laptop and the vendor.

How the same work runs on Kortix

Everything that decides how an Eigent install behaves lives on that machine. A Kortix project works the other way: the manifest and the agent folders are a repository, so the setup can be grepped, diffed and rolled back, and a new colleague reads it instead of inheriting it (Kortix docs). Session work reaches main through a change request. Merge is default-deny for agents.

The move is a checklist rather than a rebuild. The file a project starts from is the starter manifest, and it is short enough to read line by line: a default_agent, and one worker agent that names prompt_file: agents/worker.md (the starter manifest). The migration guide on this site then sets the order: list what the setup does today, write the instructions as files, wire the connectors, schedule the recurring work, run a first session on one real task, then choose where it runs.

What it costs to start is the same shape as the alternatives: Kortix Free is $0 with 500 credits a month, every feature and one project, on your own provider keys or ChatGPT subscription, with managed models needing the Credits plan or Enterprise, and paid credits start at $50 a month with members free on one pool. The difference is what the money buys. The account holds credits and people; the repository holds the company.