Sweep

You will like Sweep for one thing: quick next-edit suggestions inside JetBrains editors, where the autocomplete field is thinner than in VS Code and a fast model that predicts your next edit rather than your next token is genuinely useful. That is the trait that decides it, and it is a narrow one; the agent beside it is an afterthought you will open twice.

Pick it for autocomplete in IntelliJ or PyCharm and nothing else. Pick Junie if you want an agent that runs tasks rather than finishing lines, and GitHub Copilot if you want completion that also works in the other editors your team uses.

The pivot is the risk. Sweep began as an open-source bot that turned GitHub issues into pull requests; the website now sells a closed JetBrains plugin on proprietary models with no model choice, so a user who came for the first product cannot buy it and a user of the second cannot audit what runs it. Anyone searching for the old product finds a different one under the same name.

A vendor that changed products once will change them again, so keep nothing configured that would hurt to lose. What it does right: Privacy Mode keeps code out of training, and it is a switch rather than a sales call.

No benchmark is claimed, and the models are described only as proprietary, so there is nothing to review on the completion side beyond a latency claim without a measurement. The gap is verification: with no terminal execution, the agent cannot run tests, and its edit loop ends at the file write, so every change is unverified by construction. Context gathering is not described.

A company whose original product was a verifiable pull request now ships one whose output is a diff. The observation: the architecture moved backward while the marketing moved forward.

An open-source bot with no revenue pivoted into a paid JetBrains plugin on proprietary models, which is the open-source-without-a-business-model story that found one by picking a niche. The plugin's entire distribution is the JetBrains marketplace, and JetBrains ships Junie inside the same IDEs, so the landlord is also the competitor. Competing with the platform owner on its own storefront is not a moat; it is a lease with a review clause.

Likely outcome: acquisition by an IDE vendor for the model, or a second pivot. Position: a personal subscription, nothing a team should depend on.

The demo is fast autocomplete in IntelliJ. Procurement has nothing to work with: Pro at $20 is $1,200 a month for sixty, with no enterprise plan, no SSO, no audit logs, no self-hosted deployment and no CI mode, because the product lives inside an editor and nowhere else. The security questionnaire has no page to point at.

Vendor stability is the larger concern: a small company, a single product, no enterprise motion. Onboarding is a plugin install, which is the only easy part. Not yet.

The LICENSE in sweepai/sweep is a Sweep Enterprise Edition License with MIT carved out for the free parts, which are the remains of the GitHub bot, not the plugin anyone pays for; the open source is the part nobody uses. No bring-your-own key, no local models, no way to see what the proprietary models do with my buffer.

What I can bend: remote MCP servers with OAuth, which is a real door, and not much else. A fork of the MIT parts gives me a bot the vendor abandoned. Nothing here is mine.