EEVEE

Agents work across office apps build simple internal apps!

Open Source

Artificial Intelligence

GitHub

Security

OpenAI Day

Visit WebsiteSee on Product HuntGithub ⧉Twitter ⧉Vercel ⧉

Hi Product Hunt — I'm Nishchay, the solo maker of EEVEE. I want agents to do more than tell me how to finish office work. I want them to build the small tools I need and use them, without handing them unchecked authority over my records. EEVEE is a WebMCP workbench built around that boundary: the agent builds a small app, runs a behavioral test suite in the browser, and brings the result to you. Your passkey publishes it. Once it's running, reads happen immediately; applet writes are rehearsed against current data so you can see exactly what will change before approving. You can reject with a reason or grant a short, limited autonomy lease. For a concrete example, the included Meridian Ops applet takes an order through allocation, shipping, invoicing, and payment. The agent works through typed tools in the same tab where you see the app and review its proposed changes. There's also a versioned file library and editors for documents, spreadsheets, slides, and PDFs. GPT-6 Astra was my build partner across the whole project: concept ideation, coding, testing, the landing page, and the demo video. The interesting part wasn't just generating code. It was working through how an agent-built app should prove itself, how a person should review a proposed write, and how to make those boundaries understandable on screen. Astra helped me develop and test that loop, then explain it through the site and demo. I supplied the direction and made the product decisions; Astra helped turn them into a working system. One distinction matters: EEVEE doesn't embed a model or its own chat. Astra helped build the workbench; a WebMCP-capable browser agent operates it. The authority boundary belongs to the app, not to a promise in a prompt. The workbench is open source under Apache-2.0. The Office editors were ported from my own earlier project; the applet, evaluation, WebMCP, and approval system were built for this project. Try the demo: have the agent prepare an applet, review its test evidence, publish with your passkey, then approve or reject a rehearsed write. A passing suite only proves the scenarios it covers, and a passkey doesn't make an agent right — it makes it ask. I'd love feedback on two things: what recurring task would you trust an agent to build a small app for, and what would you need to see before approving its first write?