This entry covers what used to be Windsurf. Cognition acquired Codeium in 2025, spent months merging the product with Devin, and completed the rebrand to Devin Desktop on the second of June 2026 through an over the air update. Windsurf dot com now redirects to devin dot ai slash desktop, legacy Cascade was retired on the first of July 2026, and the credit system was replaced by daily and weekly quotas. I lead with that history because it is the single most important thing to know before you try to learn this tool.
The documentation itself is solid product writing. Installation across Mac, Windows and Linux, an onboarding flow that imports VS Code or Cursor settings, and then the actual substance: Devin Local, described as the next generation agent harness shared with the Devin CLI, running on your machine as the primary local agent. The docs cover the agent panel, terminal integration, MCP servers for extending agent capabilities, memories and customisation rules, codebase context, workflow automation, one click deployment, and remote server and SSH connections. There is also uninstallation guidance and a note about incompatibility with certain AI extensions, which is the kind of unglamorous detail that saves people an afternoon.
The memories and rules documentation is the part I would read most carefully. Getting an agent to behave consistently in your codebase depends almost entirely on what persistent context you give it, and the docs treat this as a real configuration surface rather than a magic feature. Same for MCP servers: if you want the agent to reach your database, your ticket tracker or your internal tooling, this is where you learn how, and the writing is practical. Agent Client Protocol support is worth understanding too, because it is the mechanism by which the editor stops being a walled garden and starts being one client among several.
Now the criticism, and most of it is about the rebrand rather than the docs. A very large amount of community knowledge about this tool was written under the Windsurf name and about the Cascade agent. YouTube tutorials, blog posts, Reddit threads, comparison articles, prompting guides, all of it references a product name that no longer exists and an agent that was retired. When you search for help you will find confident answers about credits that no longer apply, about Cascade behaviour that has been replaced, and about a UI that has changed.
For a learner this is genuinely costly. Product rebrands are a business decision and not my concern, but the effect on the learning path is real and it is why this sits at three and a half rather than four. My working rule is to trust the official docs and treat anything dated before June 2026 as historical. The pricing change compounds this.
The old model gave Pro users a monthly credit allowance. The current model uses daily and weekly quotas with overage billed at API pricing rather than sold as credit packs. Every cost comparison article written in the credit era is now misleading, and there are a lot of them. The free tier is decent, with unlimited tab autocomplete and light agent usage that supports maybe two or three days of active work, which is enough to form a real opinion before paying.
The second criticism is about scope. These are docs for a product, not training in how to work with a coding agent. There is nothing on the discipline that actually matters: how to review a diff an agent produced, how to scope a task so the agent does not wander, when to stop and write the code yourself, how to keep an agent from quietly rewriting something it should not have touched. That skill set is the difference between an agent that speeds you up and one that generates work for you later, and you will have to learn it from practice or from other people's writing.
On who should use this. If you already work in VS Code or Cursor, the settings import means trying Devin Desktop costs you an evening. The agent is good and the local execution model has real advantages over cloud only alternatives. If you are new to AI assisted coding entirely, I would still start with Cursor or Claude Code purely because the volume of current community material is larger and less confusing.
My three point five is for good documentation of a good tool in a bad state of transition. In a year, once the community catalogue has caught up with the name, this probably rates higher.