Test PCRC on a spare computer.
Use a disposable workspace and synthetic files. A successful tool response alone does not prove the file, image or screen result is correct: inspect the result too.
1. Install and connect
npx --yes pcremotecommander@beta setup
If npx is not recognized, run the first-install fallback once. It saves Node.js/npm/npx for your user account. Close all terminal windows and reopen one before testing the short command again.
In your PUBLIC browser, sign in, compare the terminal and website codes, review permissions and approve. The terminal should show Online and the correct account; the dashboard should show the computer selected. Keep the terminal open.
Stop with Ctrl+C, then run the same command again: the saved device should reconnect without another pairing unless revoked or expired. Briefly disconnect the test computer’s network: it should show Offline, then reconnect when restored. Never retry an uncertain write until you inspect whether it completed.
2. Connect your assistant
Use OAuth with your own account and the remote endpoint below. New connections request pcrc:read and pcrc:control. Confirm the selected computer belongs to that account.
https://pc-remote-commander.pcremotecommander.workers.dev/mcp
Use PCRC get_connection_status and get_capabilities. Report the connected computer, approved workspace and enabled capabilities. Do not change anything.
Repeat in ChatGPT, Claude and Cursor separately. For local desktop MCP use the local configuration guide. Record which client and version actually passed; one client’s success does not verify another.
3. Check files, code and recovery
Create a NEW pcrc-test.txt in the approved workspace containing PCRC_TEST_001. Do not overwrite an existing file. Read it back, inspect its SHA-256 and search for PCRC_TEST_001. Preview replacing it with PCRC_TEST_002. Wait for my review before applying the edit. After approval, verify the content, then restore its recovery backup and verify PCRC_TEST_001 again.
Repeat with small synthetic .py, .js, .ts, .json, .html, .css and .md files. These are UTF-8 text operations; changing code does not mean it was executed or validated. Test bulk move/delete with dry run first, only on disposable files, and inspect every per-file result. A changed SHA-256 must reject a stale edit. Paths outside the approved file workspace must be denied.
4. Check commands, browser and desktop
Run only this Windows command after approval: cmd /c echo PCRC_COMMAND_TEST. Verify its output is PCRC_COMMAND_TEST.
Then use PCRC to open example.com in its supported browser, inspect the page and capture it. For desktop capture, first hide private windows, capture a fresh frame and ask the assistant to describe visible synthetic text. The assistant must receive and inspect an image, not just report a successful capture. For website-approved captures, retrieve the completed image with get_action_result if instructed. A size-limit warning is a failure to deliver the image, not a visual pass.
Finally ask to open Notepad and type a harmless marker into a new unsaved document, after reviewing the action. Verify visually. Closing the test window must not discard unrelated work. Unsupported desktop applications or missing browser components should produce an actionable error.
Visual control: inspect, act, inspect again
Use PCRC visual control on my disposable test page. Inspect the page with inspect_browser_page, locate the test button, and ask for approval before interact_browser clicks it. Capture the page again with capture_browser and verify the visible change. Stop if the target is ambiguous or an error occurs. Do not enter passwords or payment details.
For the physical desktop cursor, use capture_desktop, inspect the image and coordinate metadata, then interact_desktop with action move, click, drag or scroll and that fresh frame_id. Capture again after each action. Each frame is one-use, expires after 30 seconds and must match the current foreground window; get a new frame if it expires. Move the physical pointer to a screen corner to trigger the local failsafe.
This is a fresh-frame visual loop, not a continuous live video stream. Each capture shares screen contents with your AI provider and counts as a tool request. No unattended or background recording is started. Stop the conversation’s action or press Ctrl+C in the PCRC terminal to disconnect.
5. Check approvals and isolation
Decline one new-file action: no file should appear. Approve a fresh action and verify it happens once. ChatGPT and Claude control their own confirmation UI and remembered approvals; PCRC cannot force a popup on every call. Other connections use the website approval flow. A spoken “approved” alone must not grant website approval.
Use a second account in a separate browser profile: it must not list, select, read or control the first account’s computer, receipts or billing. Revoke the disposable device: requests must stop working, and restarting it must require fresh pairing. Check recent activity and usage without exposing tokens or file contents.
Multimedia coverage: what is ready?
| Feature | Current test target |
|---|---|
| Text and code | Read, create, preview/edit and recovery; bounded UTF-8 files. |
| Browser / desktop screenshots | MCP image delivery and actual assistant visual inspection. |
| Image-file cropping / transformation | Dedicated tools not implemented yet. |
| PDF extraction / creation | Dedicated tools not implemented yet. |
| Audio transcription / editing | Dedicated tools not implemented yet. |
| Video frames / trimming | Dedicated tools not implemented yet. |
| Arbitrary binary media transfers | Dedicated bounded transfer tools not implemented yet. |
Installed external software may be used through explicitly approved commands, but that does not establish built-in support for every media format. Do not mark pending features as passed.
Record the result
For each check record client, OS, PCRC version, expected result, actual result and pass/fail. Test sandbox checkout, signed webhook activation, cancellation and billing history separately; do not make a real payment just to check the connection. A complete release also needs clean installation, update/removal, account deletion, restore and rollback evidence.
When reporting failure, include the tool name and error text. Remove tokens, pairing codes, authorization URLs and private screen/file contents. Get help.