Transcription
Today I'm giving you a full tour of Devon desktop.
The work of software engineering is shifting toward agent management. The best engineers aren't pair programming with one agent at a time. They're scoping and planning work, delegating to cloud agents, reviewing progress, and deciding what ships. Devon desktop is built for that role. One surface to manage local and cloud agents, your pull requests, and your context with the full editor right there when you still need to jump in yourself. Let's walk through it.
Here's the heart of Devon desktop. The agent command center. The default surface for managing your agents. Every agent local and cloud plus pull requests and context all in one view.
Work is organized into spaces. A space groups everything for one task. Agent sessions, PRs, files, and shared context. So related agents can actually collaborate. New sessions in a space inherit its context. So I never have to reexplain the project. Switching spaces is just switching tasks except each task has a team of agents inside of it.
This is Devon local, our first generation agent harness, the same one that powers Devon CLI. It runs right on my machine with access to my files, tools, and environment. A few things that make it special. It's built for today's models about 30% more efficient than our previous agent. It can spawn sub agents, independent workers that handle subtasks in the foreground or background while the main agent keeps going. And you stay in control with fine grain permissions. Allow, ask or deny roles scoped to file reads, writes, commands, and network access. Plus optional OS level sandboxing.
Devon desktop supports all the major models, the latest from Anthropic, OpenAI, and Google. Our own SWE family, plus leading open models like Kimmy, GLM, and DeepSeek. Pick adaptive in the model picker, and you don't have to think about models at all. It analyzes each prompt and routes it to the best model for the job. Fast and efficient ones for simple tasks and more complex ones for complex work, so your usage allowance lasts longer.
When it's done, quick review gives me a dedicated sub agent for instant feedback on the changes before I commit.
Now my favorite part, deep codebase understanding. AI that writes code is great, but the bigger superpower is AI that helps you understand code. Devon desktop attacks this at two levels. Deep wiki is symbol level understanding. Hover over any function, class, or variable. Command shift-click and instead of a bare type signature, you get a real explanation what it does, how it's used, how it fits into the codebase. And if I want to act on it, one click sends that explanation straight to my agent conversation as context.
Code maps is system level understanding. I give it a question. How does authentication work? And a specialized agent explores the repo and builds a hierarchical map showing execution order and how components relate. Every node is clickable and jumps me straight to that file and function. And they're sharable. Send a teammate a link and they can explore the map in their browser. Perfect for onboarding or design discussions.
And Devon desktop is built to be open just like Devon uses the best models for each task instead of being tied to one. The command center isn't tied to our agents through the agent client protocol ACP, which is an open standard that does for coding agents what LSP did for language servers. I can plug into codec CLI agent open code or my own custom agent and run them side by side with Devon local managed from the same interface. Use the agents that work best for each part of your development life cycle.
And when the task is too big to babysit, I delegate it. With one click, I send the plan to Devon, our autonomous cloud agent. It spins up its own VM with a desktop and browser, works a task end to end, coding, testing, even using a browser, and opens a PR I can review right here in the editor. I can keep coding locally or close my laptop entirely.
And by the way, everything that you love about the editor is still there. As you type, you still get contextually aware inline and multi-line completions. And supercomplete still predicts your intent, suggesting edits even when they're not at your cursor, like fixing a rename everywhere it's used. Hit tab and keep moving. That hasn't gone away.
So, that's Devon desktop. Grab it at devonai.top and check out the docs at docs.devonai.top to get the most out of it.