{"id":4329,"date":"2026-11-02T23:02:24","date_gmt":"2026-11-02T21:02:24","guid":{"rendered":"https:\/\/success-web.site\/?post_type=sws_case&p=4329"},"modified":"2026-11-03T13:16:37","modified_gmt":"2026-11-03T11:16:37","slug":"workshop-inventory-ai","status":"publish","type":"sws_case","link":"https:\/\/success-web.site\/en\/cases\/workshop-inventory-ai\/","title":{"rendered":"Workshop: an inventory app with AI agents, search by meaning and an MCP server"},"content":{"rendered":"\n\n
\n\t
\n\t\t
\n\t\t\t
\n\t\t\t\t
\n\t\t\t\t\tBefore<\/span>\n\t\t\t\t\t

More and more parts, and only one memory<\/h2>\n\t\t\t\t<\/div>\n\t\t\t\t

The workshop builds prototypes: boards, sensors, motors, batteries, 3D printing, soldering. To start a new build you need to know what is already on the shelf, how much of it there is, where the datasheet lives and whether the equipment can handle the job. All of this used to be kept in our heads and on the shelves: to find out whether we had the right board, someone had to go and look.<\/p>\n

The first system was planned too big: 17 models and 18 tables, and it never got as far as an inventory interface, because nobody used it. So we started again with a different method: the smallest working version, a trial in real work, conclusions, then growth.<\/p>\n<\/div>\n\t\t\t<\/div>\n\t\t\t

\n\t\t\t\t
\n\t\t\t\t\t<\/g><\/svg><\/span>\n\t\t\t\t\t

Entering things by hand takes too long. If every board has to be described in a form, the records fall behind the shelf and the inventory stops being true.<\/p>\n\t\t\t\t<\/div>\n\t\t\t\t

\n\t\t\t\t\t<\/g><\/svg><\/span>\n\t\t\t\t\t

A name doesn\u2019t tell you the job. You look for \u201cwhat can drive servos\u201d, and the table says PCA9685. Keyword search won\u2019t find that.<\/p>\n\t\t\t\t<\/div>\n\t\t\t\t

\n\t\t\t\t\t<\/g><\/svg><\/span>\n\t\t\t\t\t

Knowledge is scattered. The datasheet, the board photo, the equipment\u2019s limits and the process video belong with the record, not in browser bookmarks.<\/p>\n\t\t\t\t<\/div>\n\t\t\t\t

\n\t\t\t\t\t<\/g><\/svg><\/span>\n\t\t\t\t\t

Agents need boundaries. They have to read and change the inventory, but with no excess permissions and no making things up about what nobody has checked.<\/p>\n\t\t\t\t<\/div>\n\t\t\t<\/div>\n\t\t<\/div>\n\t<\/div>\n<\/section>\n\n\n

\n\t
\n\n\t\t
\n\t\t\tSolution<\/span>\n\t\t\t

What we did<\/h2>\n\t\t\t

A Laravel and React app with its own MCP server: people, the chat and external agents all work with the same data and tools.<\/p>\n\t\t<\/div>\n\n\t\t

\n\t\t\t
\n\t\t\t\t
\n\t\t\t\t\t

An inventory where every part has a code, documents and photos<\/h3>\n\t\t\t\t\t

Every part gets an internal code such as PRT-000141, and both people and agents find it by that code. Quantity is kept with units, the status follows it, and the type, domains and tags come from vocabularies that prevent duplicates: LEDs and LED are one term, while ESP32-S2 and ESP32-S3 are different ones.<\/p>\n

Photos, datasheets, links and notes attach to the part. Equipment records describe what the workshop can do and where its limits lie. A project brings together parts, equipment, videos and 3D models you can rotate right on the page. Videos of up to 5 GB live on Google Drive and play inside the app without a public link.<\/p>\n<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t

\n\"Parts\t\t\t\t<\/div>\n\t\t\t<\/div>\n\t\t\t
\n\t\t\t\t
\n\t\t\t\t\t

Search by meaning, not by name<\/h3>\n\t\t\t\t\t

Describe the task in your own words and the app finds what you need. The query \u201cprotect a lithium battery\u201d brings up BMS protection boards, followed by cells and charging modules. The Match column shows how close a result is to the query. A description can be in one language and the query in another.<\/p>\n

Under the hood: text-embedding-3-large at 1536 dimensions and pgvector inside PostgreSQL. The vector covers the name, part number, type, domains, tags and description, along with documents and AI descriptions of photos, so a hit in a datasheet lifts the part it belongs to. The code, status and quantity deliberately stay out of the vector: a stock change recalculates nothing. Every 10 minutes the app compares the SHA-256 of each text and updates only what has changed.<\/p>\n<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t

\n\"Search\t\t\t\t<\/div>\n\t\t\t<\/div>\n\t\t\t
\n\t\t\t\t
\n\t\t\t\t\t

A voice note becomes a part in stock<\/h3>\n\t\t\t\t\t

Forms alone will never fill an inventory, so everything goes to the inbox: voice, a photo, a file, a link, a note or a video. Speech becomes text with hints from the workshop\u2019s vocabulary, AI describes every photo, and an agent takes the item, locks it for 30 minutes and creates the records.<\/p>\n

A real example. On 28 October at 15:57, a 26-second voice note about a PWM board, asking to find the datasheet. By 16:07 the item was closed: part PRT-000141 existed, the agent had worked out the quantity from the photo, the official 52-page NXP datasheet was attached and the specs were in the description. In a separate paragraph the agent honestly wrote what it had not checked: the boards were never powered on, and an unbranded chip on the board might be a clone.<\/p>\n<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t

\n\"Inbox\t\t\t\t<\/div>\n\t\t\t<\/div>\n\t\t\t
\n\t\t\t\t
\n\t\t\t\t\t

A chat inside the app and an MCP server for external agents<\/h3>\n\t\t\t\t\t

In the chat you can ask: \u201cI want to drive 16 servos from one ESP32, what do we have in stock?\u201d The agent runs several searches by itself, opens the cards and answers with a table of codes: what to take, what is missing, what to watch in the power supply. Every tool call is visible in the answer. Before the first message you pick the mode: read only, or read and write. Deleting isn\u2019t possible in the chat.<\/p>\n

The same is available from outside: our own MCP server with 30 tools, Streamable HTTP, OAuth 2.1 with PKCE and dynamic client registration. The owner chooses read, write and delete permissions on the consent screen, and an agent doesn\u2019t even see tools outside those permissions. Tested with Claude Code and the claude.ai connector.<\/p>\n<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t

\n\"Chat\t\t\t\t<\/div>\n\t\t\t<\/div>\n\t\t<\/div>\n\n\t<\/div>\n<\/section>\n\n\n
\n\t
\n\t\t
\n\t\t\tStack<\/span>\n\t\t\t

What it is built with<\/h2>\n\t\t\t

Laravel and React with no page builders. Vectors live in the same database as the data, and the chat and the MCP server share the same tool classes.<\/p>\n\t\t<\/div>\n\t\t

\n\t\t\t
\n\t\t\t\t
    \n\t\t\t\t\t
  • <\/g><\/svg><\/span>Laravel<\/span><\/li>\n\t\t\t\t\t
  • <\/g><\/svg><\/span>PHP<\/span><\/li>\n\t\t\t\t\t
  • <\/svg><\/span>Inertia.js<\/span><\/li>\n\t\t\t\t\t
  • <\/g><\/svg><\/span>React<\/span><\/li>\n\t\t\t\t\t
  • <\/svg><\/span>TypeScript<\/span><\/li>\n\t\t\t\t\t
  • <\/svg><\/span>PostgreSQL<\/span><\/li>\n\t\t\t\t\t
  • <\/g><\/svg><\/span>pgvector<\/span><\/li>\n\t\t\t\t\t
  • <\/g><\/svg><\/span>OpenAI API<\/span><\/li>\n\t\t\t\t\t
  • <\/g><\/svg><\/span>MCP<\/span><\/li>\n\t\t\t\t\t
  • <\/g><\/svg><\/span>Google Drive<\/span><\/li>\n\t\t\t\t<\/ul>\n\t\t\t<\/div>\n\t\t<\/div>\n\t<\/div>\n<\/section>\n\n\n
    \n\t
    \n\t\t
    \n\t\t\t
    \n\t\t\t\t
    \n\t\t\t\t\tResult<\/span>\n\t\t\t\t\t

    What came out of it<\/h2>\n\t\t\t\t\t

    Numbers from the app\u2019s dashboard as of 2 November 2026. The project is ours, we use it ourselves, and it keeps growing.<\/p>\n\t\t\t\t<\/div>\n\t\t\t<\/div>\n\t\t\t

    \n\t\t\t\t
      \n\t\t\t\t\t
    • \n\t\t\t\t\t\t<\/g><\/svg><\/span>\n\t\t\t\t\t\t
      \n\t\t\t\t\t\t\t

      141 parts and 7 pieces of equipment<\/h3>\n\t\t\t\t\t\t\t

      The inventory is filled in: every part has a code and a status, and 140 are in stock right now. All the equipment works, and you can see what the workshop is capable of.<\/p>\n\t\t\t\t\t\t<\/div>\n\t\t\t\t\t<\/li>\n\t\t\t\t\t

    • \n\t\t\t\t\t\t<\/g><\/svg><\/span>\n\t\t\t\t\t\t
      \n\t\t\t\t\t\t\t

      42 inbox items processed<\/h3>\n\t\t\t\t\t\t\t

      Voice notes, photos and files became inventory records. There are no active items right now: nothing is waiting in the queue.<\/p>\n\t\t\t\t\t\t<\/div>\n\t\t\t\t\t<\/li>\n\t\t\t\t\t

    • \n\t\t\t\t\t\t<\/g><\/svg><\/span>\n\t\t\t\t\t\t
      \n\t\t\t\t\t\t\t

      235 photos described by AI<\/h3>\n\t\t\t\t\t\t\t

      All 235 out of 235. The descriptions go into the vectors, so photos work for search too.<\/p>\n\t\t\t\t\t\t<\/div>\n\t\t\t\t\t<\/li>\n\t\t\t\t\t

    • \n\t\t\t\t\t\t<\/g><\/svg><\/span>\n\t\t\t\t\t\t
      \n\t\t\t\t\t\t\t

      86 changes by agents in 7 days<\/h3>\n\t\t\t\t\t\t\t

      Agents really do work with the data. Background jobs had zero failures over the same period.<\/p>\n\t\t\t\t\t\t<\/div>\n\t\t\t\t\t<\/li>\n\t\t\t\t\t

    • \n\t\t\t\t\t\t<\/g><\/svg><\/span>\n\t\t\t\t\t\t
      \n\t\t\t\t\t\t\t

      877 automated tests and PHPStan level 7<\/h3>\n\t\t\t\t\t\t\t

      22,864 lines of PHP in the app and 16,484 lines of tests. From the first commit to the current version: 38 days and 126 commits.<\/p>\n\t\t\t\t\t\t<\/div>\n\t\t\t\t\t<\/li>\n\t\t\t\t<\/ul>\n\t\t\t<\/div>\n\t\t<\/div>\n\t<\/div>\n<\/section>\n\n\n

      \n\t
      \n\t\t
      \n\t\t\t