MakerLAB Tools

An equipment catalogue and AI assistant for a makerspace. Students find a machine, read how to use it, and ask an assistant that answers from the lab's own inventory and manuals, citing the page. Staff keep the inventory, work maintenance tickets and add new equipment in the same app. Everything is also served over MCP, so people can use the lab from Claude, ChatGPT or any other AI client.

Live at the Cornell Tech MakerLAB: https://www.makerlabai.app · Run it for your own lab: docs/fork.md

---

Where it comes from

MakerLAB Tools was built for, and runs at, the Cornell Tech MakerLAB, the fabrication lab on Cornell's New York City campus. It catalogues the lab's ~100 machines (laser cutters, 3D printers, CNC routers, electronics benches and hand tools), searches about 50 of their manuals, and is what students reach when they scan the QR label on a machine.

The software is not tied to Cornell. Everything specific to the lab (names, logos, photos, the floor plan, the sign-in domain) lives in data and environment variables, so another makerspace can fork it and run its own copy. With no configuration at all it starts on demo data.

Paper. The system was presented as a demo at the International Symposium on Academic Makerspaces (ISAM 2026), Houston: MakerLAB AI: Tools to Operate, Fix, and Build in Makerspaces (Steinberg, Parikh and Ramirez Peraza). The abstract, figures and PDF are in docs/isam-2026-demo/.

---

Screens

Catalogue. Fuzzy search and facets (status, category, material, location) across every machine.	The assistant. Opened on a tool page, it answers from that machine's manuals and cites the pages.
Reporting a problem. It troubleshoots first, then files a ticket against the right machine.	Floor map. Where each machine is, zone by zone.
Intake. Staff add equipment from a name, a link or a photo; research agents draft the record for approval.	Maintenance. Staff work the ticket queue; changes the assistant proposes wait for Confirm.
Kiosk. A lab status board for the TV at the door, with a QR code to the assistant.	MCP. Connect Claude, ChatGPT or Codex to the lab's catalogue.

More, with a walkthrough video, on the product page: https://www.makerlabai.app/product.

What it does

• Tool pages. Specs, location, required training and PPE, manuals and SOPs, and a live table of individual units with status and condition. Printed QR labels open the right page.
• An assistant on every page. Opened from a tool page it becomes a specialist on that machine: it searches the machine's manuals and cites the page, reads photos of error screens, replies in the language it is asked in, troubleshoots first and files a maintenance ticket when something is genuinely broken.
• Research and intake. Staff add equipment by pasting a list or a product link, or by photographing a label. Background agents identify each item and research its name, description, manuals and a product photo, then leave a draft for a person to approve.
• Admin. /admin holds the inventory editor, maintenance, corrections and proposal queues, project moderation, user roles and an optional one-way mirror into Notion. Anything the assistant or a student creates starts unpublished.
• MCP server. The catalogue is a Model Context Protocol server at /api/mcp: read-only without an account, or acting as yourself after signing in. See docs/mcp.md.
• Twelve languages, white-label. Interface strings go through next-intl; branding, colours, institution and assistant name come from environment variables.

---

Quick start

git clone https://github.com/philosophercode/makerlab-tools.git
cd makerlab-tools
npm install                  # Node 24 (.nvmrc)
npm run dev                  # http://localhost:3000

It runs with no configuration. With DATABASE_URL unset the app uses an in-process Postgres (PGlite) seeded with demo tools, shows a demo data banner, and stores uploads in .blob-data/. Add AI_GATEWAY_API_KEY to .env.local to turn the assistant on. .env.example documents every variable, and docs/fork.md walks through rebranding and hosting your own copy.

Feature	Needs
Tool pages, search, facets, QR labels, twelve languages, kiosk	Nothing; works on the demo data
The assistant, answering from your manuals with page citations	AI_GATEWAY_API_KEY
Adding equipment from a list, a link or a photo	The assistant's key
Maintenance tickets and staff email	Sign-in; email needs Resend
MCP for Claude, ChatGPT and Codex (docs/mcp.md)	Nothing for public reads; sign-in to act
A floor map of your lab (docs/map.md)	A plan image, traced with the prompt in the guide

Commands

npm run dev            # dev server
npm run build          # db:migrate, then next build
npm run lint           # eslint
npm run typecheck      # tsc --noEmit
npm test               # vitest: unit + integration + component
npm run test:e2e       # playwright (once: npx playwright install chromium)
npm run test:all       # all of the above; must pass before a merge
npm run eval:all       # agent evals; real, paid model calls, never in CI

The test suite needs no credentials and makes no network calls. Working against a persistent local database is covered in docs/deploy.md.

---

How it's built

Next.js 16 (App Router, React Server Components), React 19, TypeScript, Tailwind CSS 4 with shadcn/ui. Postgres (Neon) through Drizzle ORM, with pgvector for manual search; Vercel Blob for files. Better Auth with Google sign-in. Models through the Vercel AI Gateway via the AI SDK; background jobs on the Workflow SDK. Vitest, React Testing Library, MSW and Playwright. Hosted on Vercel.

flowchart LR
  subgraph People
    S[Students and visitors]
    T[Staff]
    A[Claude, ChatGPT, Codex…]
  end
  S --> P[Pages: tools, map, kiosk]
  S --> C["/api/chat"]
  T --> AD["/admin: server actions"]
  A --> M["/api/mcp"]
  C --> R[Capability registry<br/>src/lib/capabilities]
  M --> R
  R --> AL[Action layer<br/>src/lib/actions]
  AD --> AL
  R --> G[Vercel AI Gateway] --> LLM[Models]
  P --> CAT[Cached catalogue<br/>src/lib/catalog.ts]
  R --> CAT
  CAT --> DB[(Postgres + pgvector<br/>PGlite locally)]
  AL --> DB
  W[Workflows: research, manuals,<br/>photos, mirror] --> DB
  W --> B[(Vercel Blob)]
  AD --> W

1. Postgres is the source of truth and the app is its admin UI. Notion is read only by the one-time import and can optionally receive a one-way copy.
2. The assistant reads live data through tool calls, so publishing a machine reaches the chat at once, with no prompt rebuild or deploy.
3. Every agent ability is declared once in the capability registry; the chat and the MCP server are thin adapters over it, behind the same permission check.
4. Slow AI work runs as durable workflows (researching a machine, finding its photo, indexing its manuals), so a page never waits on a model.

Manuals are stored as text, split into passages, embedded and reranked, so an answer names the page it came from. Everything a model writes about the catalogue is a draft or a proposal a person approves. docs/architecture-guide.md explains each decision.

---

Documentation

Document	For
docs/fork.md	Forking it for your lab. What to rebrand and replace, then where to host it.
docs/deploy.md	Local setup stages, then a Vercel deployment step by step.
docs/handover.md	Running it: accounts, routine tasks, backups, what to do when it breaks.
docs/architecture-guide.md	How it works. Start here if you are inheriting the code.
docs/assistant.md	What the assistant can do, by role.
docs/mcp.md	Connecting Claude, ChatGPT, Codex and other MCP clients.
docs/map.md	Your lab's floor map, traced from a plan image.
docs/operations.md	Monitoring, backups and restore.
docs/testing.md	Test suite runbook.
,

Contributing

.github/CONTRIBUTING.md has the setup and the checklist. Every feature starts with a spec that merges before the implementation (docs/constitution.md, Article 1), and npm run test:all must pass before any merge. Outside contributors sign the Contributor License Agreement once. Report security problems privately, as .github/SECURITY.md describes.

Credits

Built by Isaac Steinberg for the Cornell Tech MakerLAB, which Niti Parikh (Director) and Luis Rodrigo Navarro (Assistant Director) run. Thanks to Miguel Ramirez Peraza, who built the first inventory search that started it. An older Airtable-backed version is preserved at the git tag v4-final.

License

GNU Affero General Public License v3.0 (AGPL-3.0-only). You may use, change and host it for any purpose; if you let others use a changed version over a network, you must offer them its source. Names, logos and photos are not covered. docs/legal/LICENSING.md explains it in plain language.

Copyright (C) 2026 Isaac Steinberg.