MCP DeskManager
DeskManager — MCP Connector
Your own Windows PC as a tool for your AI assistant: files, terminal, browser, desktop applications and email — running on your machine, under your account, with confirmations for anything that cannot be undone.
| Connector type | Remote MCP server (Streamable HTTP) — one private URL per PC, generated by the DeskManager app |
|---|---|
| Authentication | OAuth 2.1 with PKCE (S256). Discovery: /.well-known/oauth-protected-resource, /.well-known/oauth-authorization-server |
| Runs on | Windows 10/11 (the user's own PC). The server component only handles licensing and updates |
| Tools | 14 tools, every one annotated with readOnlyHint / destructiveHint |
| Support | [email protected] |
| Privacy | protime.it/privacy |
1. How it works
DeskManager installs on a PC and starts a local MCP server. It opens an encrypted outbound tunnel (Cloudflare) so that the PC is reachable at a private hostname such as https://mcp-xxxxxxxxxxxxxxxxxxxx.protime.it/mcp without opening any router port. The identifier is 20 random lower-case letters and digits (about 1031 combinations), generated by the app, listed nowhere and never reused. An MCP client (Claude, or any OAuth-capable client) connects to that URL, completes the OAuth flow on the user's own machine, and from then on calls tools that execute on that PC. Protime's servers never see tool inputs or outputs.
2. Connecting in five steps
- Install DeskManager on the Windows PC (installer from your licence email; it self-installs to
%APPDATA%\DeskManager). - Open the DeskManager dashboard → Connection. On first start, sign in with Google and add the licence key. The tunnel registers itself and the page shows the PC's MCP URL and its Connection Keys.
- In Claude: Customize → Connectors → + Add → Add custom connector, paste the URL.
- Claude opens the authorisation page served by that PC. Paste a Connection Key from the dashboard and confirm. Done — the connection persists across conversations.
- In any chat, call
help()once: it returns the overview and, from then on, every wrong command answers with the exact help call to fix it.
3. Tools
Read tools and write tools are deliberately split, so a client can allow the first group freely and gate the second.
| Tool | What it does | Hint |
|---|---|---|
help | Reference guide; help(section=…, query=…) returns only the entries you need | read-only |
system | Sessions, PC status, screenshots, health, clipboard | read-only |
terminal_reader / terminal_executor | Read output of persistent shell tabs, or wait server-side until a word appears, a port or URL answers, a file exists or a process ends (at most 4 minutes per call) / run commands and scripts (PowerShell, cmd, WSL). Bash scripts run from a file, so variables and quotes reach bash exactly as written | read-only / destructive |
file_reader / file_writer | Read, list, search, diff any file type (PDF, DOCX, XLSX, images…) / write, edit, move, trash, download | read-only / destructive |
browser_reader / browser_action | Built-in browser (Playwright) or, with the optional extension, the user's own Chrome: read pages, elements, tables; click, type, fill forms, wait, dismiss dialogs. Every action also reports what it changed, in the same call (observe): the page, the focused field and its real value, dialogs and controls that appeared, and whether a click opened a new tab or left the page unchanged. In the user's Chrome, clicks, key presses and mouse hovers are delivered as real input through Chrome's debugger channel — Chrome shows its yellow "DeskManager is debugging this browser" bar for the duration of each action, and the user can stop it from there. Logging in uses the passwords Chrome already stores, or "Sign in with Google": the assistant only sees account names, Chrome fills the password, and every login requires the user's confirmation. Chrome's own dialogs (password warnings, permission prompts) are detected and can be closed. The assistant stays on the tab it attached to and never follows the user's focus | read-only / destructive |
software_reader / software_action | Desktop UI automation via Windows UI Automation: read windows and controls / click, type, launch | read-only / destructive |
email_reader / email_writer | IMAP/SMTP accounts the user configures locally: read, search across one or every mailbox in a single call / send, reply, move — sending always happens in two steps (draft, then explicit confirmation) | read-only / destructive |
phone | The user's paired Android phone: send a notification, a file or a text, ask a question, set a timer, watch a terminal tab for a keyword | destructive |
panel | An interactive card inside the chat (MCP App, also in the mobile apps) whose HTML the assistant writes for the task at hand: a file browser, a photo gallery, a chart on a file's data, a form, a system dashboard, an upload or download button. Nothing is pre-built: the card talks to the PC through a small API (files in the allowed folders, upload from the phone, one-time download links, screenshot, system status), and every tap is logged on the PC. Terminal, browser and apps are not reachable from a card: it asks the assistant instead | destructive |
4. Safety model
- Confirmation gates. Permanent deletion, process kill, elevated commands and unrecognised commands require
confirmed=true; permanent deletion, elevated commands, executable files and every login in the user's browser additionally ask the user through the dashboard or the paired phone, with the reason the assistant gives. The user has 45 seconds to answer; the buttons stay active for three more minutes, after which the request must be repeated. - Passwords never reach the assistant. Password fields are reported as "filled, N characters"; stored passwords are chosen by account name and typed by Chrome.
- Recycle Bin by default.
trashis free;deleteis gated. - Command filter. A configurable security level blocks destructive shell commands (format, diskpart, recursive delete…) and flags obfuscation patterns in scripts.
- The user's browser is opt-in. Access to the real Chrome requires a separate extension the user installs; the AI is instructed to ask before using it.
- Panels act on files only. A card in the chat can read and write files in the allowed folders, never overwrites or deletes (previous versions go to the Recycle Bin), cannot upload, write or move programs or scripts, and cannot reach the terminal or the browser. Downloads use one-time links that expire after ten minutes.
- Everything is local. Tool inputs and outputs travel only between the MCP client and the user's PC.
5. Example session
help() → overview + how to open a session system(action='session', command='new') → session_id file_reader(session_id=SID, action='list', path='C:\Users\Test\Documents') browser_action(session_id=SID, action='open', url='https://example.com') browser_reader(session_id=SID, action='extract_table') terminal_executor(session_id=SID, command='Get-Process | Sort CPU -desc | Select -First 5')
6. Testing the connector (for reviewers)
7. Endpoints
| Path | Purpose |
|---|---|
POST /mcp | MCP endpoint (Streamable HTTP). Unauthenticated requests receive 401 with WWW-Authenticate: Bearer resource_metadata=… |
GET /.well-known/oauth-protected-resource | RFC 9728 resource metadata |
GET /.well-known/oauth-authorization-server | RFC 8414 authorisation server metadata (PKCE S256, authorization_code) |
GET /mcp/authorize, POST /mcp/token | OAuth authorisation and token endpoints, served by the user's PC |
GET /health | Liveness |
GET /api/panel/download/<token> | One-time download link created by a panel (expires after ten minutes; the token is the only key) |
8. Support and reporting
Email [email protected]. Security reports are read first; please include the DeskManager build number shown in the dashboard.