SWE-grep

What is SWE-grep?

SWE-grep and the faster SWE-grep-mini are Cognition's models for one job: finding the code a coding agent needs. They power Fast Context in Devin Desktop, the new name for Windsurf, which starts on its own when the Cascade agent has to search your code. Each run makes up to 8 grep, read or glob calls at once over at most 4 turns, then returns files and line ranges rather than a summary.

Cognition says this matches frontier models at finding code while running an order of magnitude faster; embedding search can miss code paths that take several hops, and step-by-step agent search is slow. It is still current, but it only works inside Devin Desktop.

Is SWE-grep still available now that Windsurf is Devin Desktop?

Yes. Windsurf became Devin Desktop, and Fast Context there still runs on SWE-grep and SWE-grep-mini.

Devin Desktop

Cognition calls it the next generation of Windsurf and says it stays fully compatible with Windsurf setups.

Fast Context

The Devin Desktop docs still name SWE-grep and SWE-grep-mini as the models behind it.

Playground

The public playground Cognition launched it with now just points back to the announcement post.

Other agents

There is no API, MCP server or download, so the only way to use it is through Cascade.

Who is SWE-grep best for?

SWE-grep is best for developers who use the Cascade agent in Devin Desktop on large codebases, where finding the right files is what slows the agent down.

• Large codebases

  It makes up to 8 searches and file reads at once over no more than 4 turns, so the agent finds the files it needs in seconds rather than after a long chain of single searches.

• Keeping the main agent focused

  It hands back files and line ranges instead of a summary, so the main model reads only the code that matters and its context stays free of noise.

• Devin Desktop and Windsurf users

  Fast Context starts on its own when a Cascade request needs code search, with no setting or command to learn.

• Trying agent search at no cost

  Fast Context is included on the free Devin Desktop plan as well as on every paid plan.

What's included with SWE-grep?

SWE-grep comes in two sizes, runs up to 8 searches at once, returns exact files and line ranges, and starts on its own inside Cascade.

Two model sizes

SWE-grep handles harder searches, and SWE-grep-mini, distilled from it, trades a little depth for over 2,800 tokens a second.

Parallel search

Up to 8 grep, glob or file-read calls in each turn, with at most 3 turns of exploring and 1 to answer.

Files and lines, not summaries

It hands the main model the exact files and line ranges, so no fast-model summary can mislead it.

Starts on its own

Fast Context runs whenever Cascade needs to search code, with nothing to switch on.

Same tools on every OS

A fixed set of grep, read and glob tools behaves the same on Mac, Windows and Linux.

How does SWE-grep work?

1. Ask Cascade

   Give the agent in Devin Desktop a task or a question about your codebase.

2. Fast Context starts

   When the request needs code search, Cascade hands the search to SWE-grep.

3. Search in parallel

   SWE-grep runs up to 8 searches and reads at a time for a few rounds.

4. Main model gets the lines

   Cascade's main model reads only the files and line ranges SWE-grep returns, then does the work.

How much does SWE-grep cost?

SWE-grep has no price of its own: it powers Fast Context in Cognition's Devin Desktop, which is included on every plan, from Free to Pro at $20 a month, Teams at $80 a month plus $40 per seat, and Max at $200 a month.

Enterprise is priced on request and adds SAML or OIDC single sign-on, admin controls and a dedicated deployment option. Paid plans come with a usage allowance that refreshes daily and weekly; past it, extra usage is billed at API prices. The page lists no yearly prices. Full pricing on devin.ai. We checked these prices on 28 Sep 2026.

How do readers rate SWE-grep?

No one has rated SWE-grep yet. Be the first.

SWE-grep pros and cons

SWE-grep finds code fast and costs nothing extra, but it only works inside Devin Desktop and its accuracy claims rest on Cognition's own unpublished tests.

What are the best SWE-grep alternatives?

Three tools on Recaply do SWE-grep's job. Repowise is open source, so you can run it on your own server.

CursorFree

Choose it when you want fast agent search in a different editor: Instant Grep indexes code on your machine and an Explore subagent runs parallel searches, with a free Hobby plan and Pro at $20 a month.

Claude CodePaid

Choose it when you work in the terminal with Claude models; its built-in Explore subagent also keeps codebase search out of the main conversation, and it comes with Claude Pro at $20 a month ($17 billed yearly).

RepowiseOpen Source

Choose it when you want codebase context for agents other than Cascade: it indexes a repo once and answers Claude Code, Cursor, Codex or any MCP client, free for public repos or $15 a month for Pro.

Questions about SWE-grep