Build an MCP connector prototype for a SaaS product over a weekend

Filently's Valentin Uhrmeister let Claude Code build an MCP server prototype for the document filing app in a weekend; after a week of hardening it shipped in Claude's connector directory with one of three tools.

Done withClaude CodeinGoogle Drive· Nov 20, 2026

What they did

Uhrmeister builds Filently, which files documents into a user's existing Google Drive folders. An MCP server had sat on the team's list for a while. On a Saturday Uhrmeister spent about fifteen minutes brainstorming the idea with Claude, and on the Sunday let Claude Code build a working prototype with little guidance and no real back and forth. The prototype had tools to list documents waiting for review, file a document into the suggested folder, suggest a name and folder, list every folder in Drive, and show a filing history. The team then spent the following week turning it into something they would let near users' contracts and IDs.

How it went

Uhrmeister calls the prototype surprisingly good. The shipped version, live in Claude's connector directory, cut the folder listing tool (too slow for big archives and it would expose every folder name) and the filing history tool (not reliable yet). The week went to an activity log that shows when Claude filed something, a page to disconnect AI apps, hard limits (no deletes, no reading document contents), security work and a privacy policy update. In-chat onboarding was dropped because files attached in Claude could not reach the server. No user numbers yet.

Worth knowing

A weekend prototype was not the same as a shippable connector: most of the effort went into permissions, auditing and cutting tools that leaked too much. Claude's directory policy also led them to remove a tool that asked users how they found the product. The post is by the founder and promotes Filently.

Try it

With Claude Code· 5 parts to fill in

Build a prototype MCP server for [my product] so AI assistants can [the two or three jobs users would ask for]. Use our existing [API or backend] and do not give the assistant any ability to delete data or read [sensitive content]. Keep the tool list small, and for each tool tell me what data it would send to the AI provider. Log every action the assistant takes so users can see it was the assistant. Stop when the server runs locally and I can call each tool from [Claude or my MCP client], then give me a list of what would still be needed before real users connect it.

Square brackets are the parts to fill in. The prompt is ours; the story is theirs.

Where they said it

Related