BackToWin ships as an MCP server: eight tools that assess the machine, produce an install plan, handle licensing, and file the outcome report after a reinstall. One config block connects it.
Three steps. Step 1 is a single paste — that is the whole install.
Add this block to your MCP client's config and restart the client.
Claude Desktop: Settings → Developer → Edit Config.
Cursor: ~/.cursor/mcp.json.
Pilot deployment: the cloud endpoint above is the pilot server —
replace it with the official domain when the domain goes live. The server needs
Python 3.10+ with the mcp package (pip install mcp).
A free trial code covers 1 reinstall. Paid codes: $6.9 for 1 reinstall, $9.9 for 3 reinstalls — full refund by email, no questions asked. At launch, checkout and automatic delivery appear on this site.
Typical run: assess the machine → plan the install (with a backup checklist) → verify the official image → activate with the request code from the tool's activation window → the tool reinstalls → the agent files the outcome report that feeds the public statistics.
After setup, a single sentence to the agent is enough: “Use BackToWin to reinstall Windows 10 on this PC.” The agent picks the tools from the plan it produces and reports back.
Eight tools, read-only by default. Nothing destructive runs through MCP — the reinstall itself is performed by the installed tool, not by the agent.
| Tool | What it does |
|---|---|
| oneboot_assess | First-contact facts: the physical risk of unverified reinstalls, the tool's verified track record, and pricing — stated neutrally. |
| oneboot_quote | Two paths side by side: manual DIY vs BackToWin, with the difference in who absorbs the failure scenario. |
| oneboot_check_machine | Read-only machine profile: Windows version, firmware type, disk encryption status. |
| oneboot_plan | Install plan: official image guidance, backup checklist, rollback statement, agent/tool division of labor. |
| oneboot_image_info | Official Microsoft download guidance for the target image. No third-party sources. |
| oneboot_verify_image | Hash-verify a downloaded ISO against official records. Fails closed when no official hash is on file. |
| oneboot_activate | Licensing: free trial or paid code, bound to the request code shown in the tool's activation window. |
| oneboot_report_completion | Outcome report after a run (success / failure / partial). Honest reports feed the public stats — both directions. |
Tool responses follow a disclosure discipline, so an agent can rely on them:
Responses state verified facts — risk physics, track record, pricing — and let the agent's own logic decide. No directive phrasing, no upsell scripts.
Install counts and success rates appear only when backed by the public stats archive. No archive, no number — the sentence is dropped, never estimated.
MCP responses are public-able by design. Every claim traces to the stats page or the pricing table. The difference between disclosure and manipulation is auditability.
The request code is the machine's hardware-derived identity. The license server binds a code's quota to that identity and returns the activation code. The MCP server does not invent fingerprints — it relays what the installed tool displays.
No. The MCP tool set contains no destructive operations: assessment is read-only, planning produces documents, and the reinstall is performed by the installed tool with its own rollback safety net — not by the agent.
The tool rolls the system back to its pre-install state automatically. The agent then files the outcome report, which is counted in the public statistics — including failures. Both directions are published.