Your computer’s tools, inside your desktop assistant.
Run PCRC directly from Claude Desktop, Cursor, or another client that supports local MCP. The client starts a process on your computer and talks to it over stdio.
Choose the connection that fits.
| Local MCP | Remote MCP | |
|---|---|---|
| How it connects | Your desktop client launches the task MCP server with Node.js and its bundled Python backend. | Your assistant connects to the hosted HTTPS endpoint. |
| Accounts and pairing | No PCRC website sign-in, pairing code, or OAuth is needed for stdio. | Sign in to PCRC, approve a computer, then authorize the client. |
| Best for | Tools on the computer running Claude Desktop or Cursor. | Web clients such as ChatGPT or Claude web reaching your selected PC. |
Local requests do not route through the PCRC cloud gateway. Content returned to your AI client may still be sent to its model provider. Local does not mean the AI conversation is offline.
Looking for remote MCP? Follow the pairing guide →1. Prepare the Windows runtime
Requires Windows, Node.js 20+, npm, and a desktop client with stdio support. The Windows beta includes Python.
- Open PowerShell and prepare the runtime without starting remote pairing:
npx --yes pcremotecommander@beta doctor - Note the installation directory printed by
doctor. The beta.11 default is under%LOCALAPPDATA%/PCRemoteCommander/preview/0.7.0-beta.15. - Create a folder for the assistant’s work, or choose an existing folder you are authorized to use:
New-Item -ItemType Directory -Force "$HOME/PCRC-Workspace"
2. Add the local server to your client
Use the example below for Windows. Replace YOUR_NAME, the runtime version if different, and the workspace path. Use the exact install directory printed by doctor. Environment placeholders such as %LOCALAPPDATA% are not expanded in this JSON.
{
"mcpServers": {
"pcrc-local": {
"command": "node",
"args": [
"C:/Users/YOUR_NAME/AppData/Local/PCRemoteCommander/preview/0.7.0-beta.15/deploy/mcp-task-server/dist/main.js"
],
"env": {
"JAN_MCP_WORKSPACE": "C:/Users/YOUR_NAME/PCRC-Workspace",
"JAN_MCP_PYTHON": "C:/Users/YOUR_NAME/AppData/Local/PCRemoteCommander/preview/0.7.0-beta.15/.venv/Scripts/python.exe",
"JAN_MCP_CONTROL": "1",
"JAN_MCP_CAPABILITIES": "files,commands,browser,desktop",
"JAN_MCP_DURATION_MINUTES": "60",
"JAN_MCP_BROWSER_CHANNEL": "msedge"
}
}
}
}Claude Desktop
- Open Claude Desktop’s app settings, then Developer → Edit Config.
- Add the
pcrc-localentry insidemcpServersinclaude_desktop_config.json. Preserve existing server entries. On Windows the file is under%APPDATA%/Claude/. - Save, fully quit Claude Desktop, and reopen it. Review and enable the server’s tools.
Official local MCP setup guide ↗
Cursor
- Open Cursor’s MCP configuration from Customize, or edit your existing
mcp.json. - Add the same
pcrc-localentry while preserving other servers. This configuration usescommandandargs, rather than a remoteurl. - Save the configuration, reload or restart the server, and enable its tools for Agent.
Official Cursor MCP configuration guide ↗
Other desktop clients
Choose a local stdio connection. Set the executable to Node.js and supply the task-server path and environment from the JSON. Match your client’s configuration format. Web-only clients cannot launch a local process on your PC; use the remote setup instead.
Linux: local source installation
The npm launcher is Windows-only. If you already have an authorized PCRC source checkout, Python 3.11+, venv, and pip, run these from that source directory:
bash setup-linux.sh bash jan-cli.sh --workspace /ABSOLUTE/approved-workspace serve
The workspace must already exist. The second command starts stdio and normally waits quietly for a client. Stop it with Ctrl+C after checking, then let your desktop client start its own process using:
{
"mcpServers": {
"pcrc-local": {
"command": "/ABSOLUTE/PCRC/.venv/bin/python",
"args": [
"-m",
"jan_mcp_commander.cli",
"--workspace",
"/ABSOLUTE/approved-workspace",
"serve"
],
"env": {
"JAN_MCP_ALLOW_WRITES": "0"
}
}
}
}Replace both absolute paths. A public Linux download and a clean Linux desktop-client test are still pending. Do not use the Windows npm command on Linux.
3. Verify before making changes
Use PCRC local to check get_capabilities and get_connection_status. Call suggest_workflow for diagnosing my computer connection. Then list the top-level entries in my approved workspace.
The Windows configuration enables file, command, browser and desktop tools for a 60-minute local lease. This is explicit local configuration, with your AI client responsible for presenting action approvals; it does not use the hosted gateway approval page. Commands run as your Windows account and desktop tools can see the screen. Test edits only on disposable files, inspect results before applying changes, and restart the server when the lease expires. Use the installed Node.js executable’s absolute path if your client cannot find node.
The Linux source example above remains workspace-scoped; a public Linux full-control distribution has not been verified.
Local MCP questions
What URL do I paste for local MCP?
There is no HTTP URL for this stdio setup. Your client starts the executable using the command and arguments. The hosted /mcp endpoint is for remote connections.
The process waits silently. Is it broken?
A stdio server waits for MCP messages from a client. Configure it in your desktop app; it does not open a webpage or interactive CLI prompt.
The server cannot start or find my folder
Run doctor again, check the exact Python path and version directory, and confirm that your approved workspace exists. Use absolute paths and valid JSON. Check your client’s MCP logs without sharing tokens or private file content.
How do I disconnect or uninstall?
Disable or remove the pcrc-local entry in your AI client and stop its running server. Keep any workspace files, backups, and private state you still need. Removing the connector does not delete those files.