Sweep (software)
Last reviewed
Jun 4, 2026
Sources
17 citations
Review status
Source-backed
Revision
v1 · 1,921 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
Jun 4, 2026
Sources
17 citations
Review status
Source-backed
Revision
v1 · 1,921 words
Add missing citations, update stale details, or suggest a clearer explanation.
Sweep (stylized with a broom emoji and reachable at sweep.dev) is an AI developer tool built by the San Francisco startup of the same name, founded in 2023 by Kevin Lu and William Zeng. It first became known in 2023 as an open-source "AI junior developer" that turned GitHub issues into pull requests automatically, drawing more than 7,000 stars on GitHub. The company later pivoted, and Sweep is now an AI coding assistant delivered as a plugin for JetBrains integrated development environments (IDEs), offering next-edit autocomplete and an in-IDE coding agent. Sweep went through Y Combinator's Summer 2023 batch and raised a small seed round backed by Goat Capital and Replit's leadership.
Sweep should not be confused with an unrelated company of the same name based in New York that builds an agentic workspace for Salesforce and HubSpot go-to-market systems; that company is a separate business with separate founders and funding.
Sweep was started in 2023 by Kevin Lu and William Zeng, who had previously worked together as engineers at the gaming platform Roblox. According to the founders, the idea grew out of the repetitive "software chores" they handled at Roblox that they believed could be automated with large language models. Zeng described the early product as "an AI-powered junior dev for software teams." The company is based in San Francisco. Press coverage and the Y Combinator profile list Lu and Zeng as the founders, while TechCrunch's reporting from November 2023 also named Samip Dahal as a co-founder. The team remained small, on the order of four people.
Sweep was accepted into Y Combinator's Summer 2023 (S23) cohort and presented at the accelerator's Demo Day that year. It launched publicly through two Y Combinator "Launch YC" posts in mid-2023: one framed Sweep as "an AI developer that handles your tech debt," and a second positioned it as an "AI-powered junior developer."
As a Y Combinator company, Sweep received the accelerator's standard early-stage investment of roughly $500,000. In November 2023, TechCrunch reported that the startup had closed an additional $2 million round at a $25 million post-money valuation. The investors named were Goat Capital, Replit chief executive Amjad Masad, Replit's vice president of AI Michele Catasta, and Exceptional Capital. Zeng said at the time that the new capital would be used to grow the team "in the coming year from two employees to five." Public databases put Sweep's total disclosed funding at roughly $2.8 million across these rounds.
| Round | Date | Amount | Valuation | Investors |
|---|---|---|---|---|
| Y Combinator (S23) | 2023 | ~$500K | n/a | Y Combinator |
| Seed | November 2023 | $2 million | $25 million post-money | Goat Capital, Amjad Masad (Replit CEO), Michele Catasta (Replit VP of AI), Exceptional Capital |
No subsequent priced rounds for sweep.dev have been publicly reported.
In its first phase, Sweep was an open-source GitHub application that behaved like an automated junior engineer. A developer installed the Sweep bot on a repository, then created or labeled a GitHub issue (tagging it "sweep"); the bot read the codebase, drafted a plan, wrote the code, and opened a pull request, responding to review comments much as a human contributor would. The example the founders frequently used was a natural-language request such as "add debug logs to my data pipeline," issued from outside the IDE. Sweep targeted well-scoped maintenance work: bug fixes, configuration changes, unit-test generation, small refactors, and similar "tech debt" tasks.
Technically, the original Sweep used OpenAI's GPT-4 (among other models) for code generation, paired with a custom code-search engine the team built for Python that combined lexical search (literal or near-literal matches) with vector search (retrieval of more loosely related code). Because generative AI is prone to errors, Sweep prompted users to review and edit any generated code before merging it. The project was released under the Apache 2.0 license and could be self-hosted via a Docker image on a laptop or a cloud provider, in addition to the hosted offering. The hosted service was priced steeply at $480 per seat per month, far above mainstream code-completion tools. The open-source repository, sweepai/sweep, accumulated roughly 7,600 to 7,700 GitHub stars, making it one of the more visible early "issue-to-PR" agents.
By 2025 and into 2026, Sweep repositioned itself. The founders concluded that the fully autonomous "AI junior developer" vision was "many years out" and chose instead to build a coding assistant developers could use day to day. They focused on JetBrains IDEs (IntelliJ IDEA, PyCharm, and the rest of the JetBrains family), arguing that existing AI assistants served JetBrains poorly compared with their Visual Studio Code counterparts, and that even GitHub Copilot's JetBrains experience lagged its VS Code version. The company's framing was that it was building, in effect, a "Cursor for JetBrains," referencing the popular AI editor Cursor.
The repository sweepai/sweep was retitled "Sweep: AI coding assistant for JetBrains," and the GitHub organization's description became "Next-edit autocomplete and coding agent plugin for JetBrains." The original README now carries a notice thanking users for their support of the old product and directing them to the JetBrains plugin. The company launched its JetBrains "next-edit" autocomplete (which it nicknamed "Cursor Tab for JetBrains") in February 2026 and reported growing from zero to tens of thousands of installs within about six months.
A notable consequence of the pivot's emphasis on autonomous coding came in March 2026, when co-founder William Zeng announced he was joining Elon Musk's xAI (and SpaceX) to work on advanced coding AI, alongside Firebender co-founder Aman Gottumukkala. Zeng framed the move as a way to "accelerate progress" toward solving autonomous engineering. His departure left the future leadership and ownership of Sweep's product less clearly documented in public sources.
Sweep is now an AI coding assistant distributed primarily as a JetBrains IDE plugin, listed on the JetBrains Marketplace as "Sweep: AI Autocomplete & Coding Agent." It supports the full JetBrains lineup, including IntelliJ IDEA, PyCharm, WebStorm, GoLand, PhpStorm, Rider, CLion, RubyMine, Android Studio, and others, under the slogan "one plugin, every IDE." The company positions speed as its main differentiator, claiming the fastest autocomplete on JetBrains and pitching itself as smarter and faster than alternatives because it taps JetBrains' own static-analysis and code-index APIs.
| Feature | Description |
|---|---|
| Next-edit autocomplete | A "Tab" model that uses the developer's recent edits as context to predict the next change and the next location to jump to, rather than only completing the current line. |
| Coding agent | An in-IDE agent that searches the codebase, edits code across files, runs tests, and checks for linter errors. |
| Inline editing | Natural-language edits applied directly to selected code. |
| AI commit messages | Generated commit messages based on staged changes. |
| AI code review | Reviews diffs between branches. |
| Codebase understanding | Indexes the project (using the JetBrains code index) so suggestions and the agent are aware of the current file and wider repository. |
| Web search and fetch | Lets the assistant pull in information from the web from within the IDE. |
Sweep advertises a custom inference stack with sub-100-millisecond autocomplete latency and, for privacy, options including local model execution, a "Privacy Mode," bring-your-own-key (BYOK) support, zero data retention, and SOC 2 compliance.
In January 2026, Sweep released Sweep Next-Edit, a 1.5-billion-parameter model that predicts a developer's next code edit. It is fine-tuned from Qwen2.5-Coder, ships under the Apache 2.0 license with an 8,192-token context window, and is distributed in GGUF format on Hugging Face (repository sweepai/sweep-next-edit-1.5B) for use with local runtimes such as llama.cpp, Ollama, and LM Studio. The company reported that the model runs locally in under 500 milliseconds using speculative decoding and outperforms models more than four times its size on next-edit benchmarks; its published table showed Sweep Next-Edit scoring about 67.8% overall versus roughly 48.3% for Qwen3-8B, 43.3% for Zed's Zeta, and 25.3% for Continue's Instinct. Sweep said it serves the model in production on a fork of TensorRT-LLM using FP8 quantization with n-gram speculative decoding. Releasing the weights, the company said, was meant to let the community build fast, privacy-preserving autocomplete for any editor; the organization also maintains experimental next-edit plugins for VS Code (vscode-nes) and Neovim (cursortab.nvim).
Sweep's JetBrains assistant uses a freemium, credit-based model rather than the flat $480-per-seat price of the original GitHub product. A free trial is available, and paid plans grant Sweep API credits equal to the plan's cost (rolling over for 30 days), with the ability to buy additional credits or enable automatic top-up.
| Plan | Approximate price | Notes |
|---|---|---|
| Free | $0 | Basic access and trial usage. |
| Pro | $20 / month | Unlimited autocomplete plus $20 in Sweep API credits for agent/chat usage; priority support. |
| Ultra | $60 / month | Higher credit allocation for heavy users. |
| Team | $40 in credits per seat | Adds SSO, usage visibility, and the ability to enforce Privacy Mode organization-wide; overage billed pay-as-you-go. |
Pricing details can change; the figures above reflect the structure published by the company in early 2026.
On the JetBrains Marketplace, the Sweep plugin holds a high user rating (around 4.9 stars) and reports more than 40,000 installs, and the company has described itself as a top-ranked AI plugin in that ecosystem. Early users have praised the autocomplete's responsiveness, with testimonials calling it dramatically faster than JetBrains' built-in tooling. Independent reviews frame Sweep's chief appeal as bringing a Cursor-quality autocomplete and agent experience to JetBrains developers who had been underserved by Copilot and similar tools. The 2023 open-source bot also earned attention in developer circles for demonstrating early issue-to-PR automation, though contemporary coverage (including TechCrunch) noted healthy skepticism about the reliability of fully autonomous code generation, the reason Sweep kept a human review step.