AI · MCP · Flutter · Android · React · Go
Partbench
An electronic parts catalog for the home lab: how many of each part there are, where they lie, what a project is missing and where to buy it. Parts are added from a photo or a description, and Claude and ChatGPT reach the catalog through an MCP server.
two71.io/projekty/partbench/en/ → (project site)The challenge
A few hundred kinds of parts in drawers, boxes and bags from TME, Botland and AliExpress. An inventory spreadsheet stops matching reality within a month, because every change means retyping. The result is always the same: a second order for a regulator that was sitting in drawer 5, and a project stopped halfway through soldering for want of one capacitor.
The solution
Each part looks like a drawer label: value, name, location, and stock changed with one tap. Stock can be counted in pieces or kept as a level (none, few, some, plenty). New parts come from a photo (one part, a handful, or a box sticker), a description, or an imported order. The model reads markings and proposes a name, category and parameters, and checks unknown markings against the TME catalogue and the web. Only what the user approves goes into the catalog. A project is a saved bill of materials from KiCad, EasyEDA or a pasted list: it shows what's missing, links to TME and Botland, and a "Built" button takes the parts off stock. An MCP server with OAuth sign-in lets you ask Claude or ChatGPT whether a project is covered and take parts off stock from the conversation.
Scope & technology
The whole product: a Flutter Android app (an APK that checks for new versions itself), a browser version in React + TypeScript, a Go API with PostgreSQL. Each catalog is isolated in the database with row-level security, and every stock change is logged with a date, reason and source (phone, browser, API, MCP). Integrations: vision models for part recognition, the TME API, web search. Value normalisation ("10k" = "10 kΩ" = "10000") is ordinary tested code, with no model involved. The MCP server exposes 14 tools and runs its own OAuth 2.1 authorisation server (PKCE, dynamic client registration) with separate read and write permissions; scripts can use API keys. The interface is in Polish and English.