Eigent

The trait that decides it is the download. This is an application you install and open, with a workforce of agents sharing context and handing results back for you to review, which puts multi-agent work in reach of colleagues who will never enjoy a command line. For a mixed team that is a real widening of who can use these tools.

Pick Claude Code instead if you live in a terminal and your work is source control, because a window adds nothing there and takes memory. This is for the other half of the office.

The exposure is the absence of a boundary. Shell access and browser control run in the user's own session, with no sandbox in the design, which means a mistaken command reaches the same filesystem as your credentials, your unpushed branches and your personal files. Several agents working in parallel multiply the chances of one of them being wrong.

Run it on a machine you would not mind rebuilding. What it does right: the workforce returns results for review rather than committing to them, so a person stands between the agents and anything final.

The design decision worth examining is the shared context. Message passing between agents duplicates information and costs tokens proportional to the number of participants; a common context avoids that and is measurably cheaper. The cost is isolation: a wrong conclusion entering the shared state is visible to every agent, and none of them has an independent view from which to contradict it.

Building on an existing research framework rather than a bespoke runtime is sound reuse. No benchmark is reported, and the decomposition strategy is undocumented.

This is a well-constructed funnel. Users on the free tier pay their own inference bill, so acquisition is nearly costless, and the paid tiers convert people who would rather not manage keys at all. The credit ladder discounts the unit as the commitment rises, which is standard and effective at pulling buyers up a tier.

What is missing is the switching cost. Nothing accumulates inside the product that a user would lose by leaving. Likely path: an enterprise deployment story, or absorption into the research ecosystem it builds on. Position: small, and watch whether anything sticky ever appears.

The paid tier at $19.99 works out near $1,200 a month at our headcount, and that is the floor rather than the bill, since the allowance is consumed by work nobody can estimate in advance. The larger cost is distribution: a desktop application on sixty machines becomes an endpoint-management obligation with updates, versions and an inventory.

Nothing runs unattended, so it contributes nothing to our pipelines, and there is no administrative console to federate identity against. Onboarding is genuinely easy, which is the one point in its favour. Not yet.

Three local serving options named explicitly is a signal that somebody on this team runs models at home. I built it from source with a clone and npm run dev, pointed it at my own server, and never touched their hosted credits or created an account.

MCP support means my existing servers are available to the agents without a shim, and Apache-2.0 means the desktop build is mine to modify and redistribute. What I would want next is a way to run it without the window, since everything underneath is perfectly scriptable and the interface is the only thing insisting otherwise.