Amazon Quick Desktop — I set it up and explored every feature. Here's everything.

Amazon Quick Desktop launched in preview in April 2026 for macOS and Windows. It extends Amazon Quick from the browser to your machine with a key architectural choice: local-first. The AI backend runs locally, your files stay on your computer, and the only network calls are to AI models via API Gateway and to your connected services (Slack, Outlook, Gmail, etc.). Your data doesn't leave your machine to be processed.

It's available to Plus and Enterprise accounts. Free accounts get a 30-day trial of the desktop app. Enterprise accounts get it across all supported AWS regions.

The MCP support is the headline feature for developers. Amazon Quick Desktop supports three ways to connect MCP servers — local servers running a command on your machine, remote servers over HTTP, and imported configurations from other tools like Kiro, Claude Code, or AIM. You can also export your config as JSON and share it across your team.

The Capabilities section is where you connect external tools. It has three tabs: Connectors, Schedules, and Tools. The MCP Servers section is the most important one — and it's more flexible than it looks.

I connected AWS MCP as one example — it gives the agent 10 tools (7 read, 3 write) for working with AWS. But this is just one MCP server. You could connect any MCP-compatible server: your own internal tools, third-party services, anything running the MCP protocol locally or remotely.

Each tool has a permission setting — Always allow or Ask each time. Read operations are set to Always allow by default; write operations default to Ask each time. That's the right default: you don't want the agent executing AWS CLI commands without a confirmation step.

This is one of the most interesting features. Amazon Quick Desktop can delegate coding tasks to external agents via ACP (Agent Client Protocol). You're not limited to Q's own coding capabilities — you can route work to whichever agent you prefer.

Under Agents & Skills → Skills, there are 26 built-in skills that ship with Amazon Quick. These are the capabilities the agent draws on when handling different types of requests.

Creating your own skills

You're not limited to the 26 built-in skills. The + Create button in the Skills panel lets you add your own — three ways:

The Schedule Management skill (17 tools) is notable — it means you can ask Amazon Quick Desktop to set up recurring tasks directly from chat, without touching a config file. The Web Browser skill (24 tools) gives Q full web interaction capability, including clicking, form filling, and reading page content.

This is what separates the desktop app from any web-based AI tool. Under My Computer → Local Folders, you add folders the agent can access and index. I added my OneDrive folder — 25.27 GB, 34,210 files across 5,957 folders.

For each folder you can enable three capabilities independently:

My Context has two tabs: Knowledge graph and Memory. These start empty but populate as you index files and use the agent. After indexing my OneDrive folder, both tabs had real data.

Knowledge graph

The knowledge graph extracts entities and relationships from your indexed files and renders them as an interactive node graph. After a partial index, it had already extracted two nodes — the folder itself and a person entity linked to it — with a directed relationship between them.

Memory

The Memory tab shows what Amazon Quick Desktop has learned and retained. After the folder analysis task, it had already stored one procedure — get_folder_size_summary(folder_path: str) — with a 70% confidence score. This is the agent learning a reusable tool from the task it just ran.

Clicking the model indicator in the chat input opens a three-option picker. The options map to different quality/speed tradeoffs — exactly what you'd expect.

The Thinking toggle (under Smart) enables extended reasoning — the model works through the problem step by step before responding. Worth turning on for anything architectural or ambiguous. Adds latency but meaningfully improves answer quality on hard questions.

I asked Amazon Quick Desktop to check my folder size. First attempt — before granting agent access to the folder — it correctly told me it didn't have access and pointed me to the right setting.

Once I enabled agent access and rescanned, it ran the analysis. The result:

The settings panel is comprehensive. Here's everything worth knowing.

Browser settings

You can set which browser the agent uses for web tasks — Auto (system default), Chrome, Edge, or Brave. The Use my browser toggle is off by default, which means the agent launches a standalone browser with no access to your accounts or sessions. Turn it on if you want Q to use your actual logged-in browser context.

The Screenshot capture shortcut (default: Ctrl+Shift+Q) lets you select a screen region and send it directly to the agent — useful for debugging UI issues or sharing error messages without typing them out.

To-do management

Amazon Quick Desktop has a built-in to-do system with four priority levels (None, FYI, Important, Urgent). You can ask Amazon Quick Desktop to create, update, or list to-dos from chat — or add them manually via the dialog. These integrate with the Activity Feed skill.

Performance & Response preferences

The Response preferences field is a global instruction that gets included with every message across all agents — like a persistent system prompt you control. Set it to "Be concise", "Always include code examples", "Respond in bullet points" — whatever fits your workflow.

The Max parallel tasks slider (1-50, default 20) controls how many concurrent extraction tasks run during batch operations like indexing large folders.

Voice — Dictation + Talkback

Voice has two modes. Dictation is speak-to-type — you talk, Q transcribes, you hit Send. Talkback is Q speaking its responses aloud, with a choice of voice, speed, and Live mode (mic stays open so you can interrupt mid-response by saying 3+ words).

There's also a Skill preloading toggle — when on, Amazon Quick Desktop classifies each query before planning and pre-activates the predicted skills. Reduces latency by front-loading skill selection.

Troubleshooting & Data

Export Diagnostics saves logs to your Desktop — useful for filing bug reports. The Danger zone clear wipes everything: conversations, cached messages, knowledge graph, saved credentials, and preferences. The app quits after cleanup. To fully uninstall, drag the app to Trash after clearing data.

A few things that stood out after going through everything:

The MCP + coding agent combination is the real value. AWS MCP gives Q direct access to your AWS environment. Coding agents (Kiro, ClaudeCode) give it the ability to write and run code. Together, they make Q a genuine automation layer — not just a chat interface on top of documentation.

Local file indexing with semantic search is underrated. The ability to ask questions over your entire local file system — with embedding-based search finding relevant content even without exact keyword matches — is genuinely useful for anyone working with large codebases, documentation sets, or project files.

The Response preferences field is worth filling in immediately. It's a global system prompt you control. Set your formatting preferences, language, verbosity level once and never think about it again across all agents.

Write operations default to Ask each time — keep it that way. The AWS MCP write tools (Call Aws, Run Script, Get Presigned Url) are all set to Ask each time by default. Don't change them to Always allow. The extra confirmation step is the right guardrail when the agent has AWS CLI access.