Windsurf (Devin Desktop)

Cascade is a solid editor agent: it edits across files, runs commands, and does not lose the plot on a medium refactor, which is the daily trait that decides it. The trouble is everything around the product. It has been sold once and renamed once inside a year, so the docs you bookmarked, the plan you bought and the name on the procurement form each point somewhere slightly different, and you will spend attention on the vendor that should go to the code.

Use it if you are already a Devin shop and the brand consolidation helps you. Otherwise pick Cursor for the editor, or Claude Code for the terminal.

The rename is the failure mode. A tool whose docs, tutorials and muscle memory point at a name that no longer resolves, windsurf.com redirects to devin.ai, has broken every link a team relied on without changing a line of code, and the settings, plans and support paths all moved with it. The isolation on offer is the Devin Local harness, with filesystem and network filtering; Cascade itself still runs your commands as you.

Wait one release cycle before trusting any document about it. What it does right: Devin Cloud agents run alongside the desktop, so a long task can leave the laptop and come back as a branch.

Devin Desktop, formerly Windsurf, documents two components worth noting. 1. Cascade edits across files and runs commands, an ordinary loop. 2. A separate planning agent pairs with it, a defensible decomposition, since plan and act loops fail differently and separating them lets each be inspected on its own. Context gathering is not described beyond the open workspace.

No benchmark is listed for the tool or for SWE-1.7, the in-house model on the list, so the one proprietary component is the one with no number. The observation: the vendor renamed the product before publishing a figure for it.

Windsurf was my example of cap-table risk, and the record confirms it. Codeium built the IDE, Cognition took the asset in July 2025, and on June 2, 2026 the product became Devin Desktop. That is not a pivot; that is an absorption, and an absorbed product's roadmap is whatever serves the acquirer's flagship, which here is Devin, not the editor.

No further acquirer; the acquisition is the story. The remaining question is whether the desktop survives as more than an on-ramp to the cloud agent. Position: use it if you are a Devin shop; otherwise, do not build a workflow on it.

The demo is Cascade refactoring with a planning agent. Any assessment done under Codeium is void: the owner is Cognition, the product is Devin Desktop, so legal starts over with a new DPA and a new questionnaire. Teams is $80 plus $40 per full seat, so sixty seats is $29,760 a year, with no headless mode for CI and no data-retention terms on the page I can quote.

Onboarding is nil for VS Code users, which is the good news. Not yet: I need Cognition's retention terms in writing and one quarter of stable naming before I put it on a form.

My checklist against the models docs: bring_your_own_model false, local_models false, open_source false. Three strikes, and a fork is out of the question. The interesting part is ACP: the IDE hosts any ACP-compatible agent, so I can treat Cascade as furniture and run my own agent in its panel, which is the one door in a closed house.

MCP servers go in a plain mcpServers config, the same shape as everyone else's, so at least my tool shelf moves in without edits. Grudging respect for the ACP hosting. Nothing for the rest, and I would not put my keys near it, since it will not take them.