# Undergrove on Product Hunt: words

Drafts for the launch listing, written to `docs/POSITIONING.md`: lead with the models and what the product does, make privacy claims only as far as the code supports them, no absolute claims. Check each against `/api/status` and the privacy page on launch day, since the models offered and the alpha’s terms can change. Product Hunt’s form counts characters; the counts below include spaces.

## Name

Undergrove

## Tagline (60 characters at most)

1. Open-weight AI, encrypted to an enclave your browser checks (59)
2. A private AI workspace on open-weight models in enclaves (56)
3. Ask open models. Check the enclave they run in. (47)

## Description (260 characters at most)

1. Undergrove is an AI workspace for open-weight models like GLM-5.3, Kimi K3 and Gemma 4. Your browser checks the attestation of the hardware enclave each request goes to, then encrypts the request to it, so our relay passes on ciphertext it can’t read. (251)

2. Web search, files, Python in your tab, model races and encrypted shared projects, on open-weight models in hardware enclaves. Each answer can show its route and the bytes that left your browser. Live models are in private alpha. (228)

3. An OpenAI-compatible API and a workspace for open-weight models in Tinfoil’s hardware enclaves. The workspace seals each request in your browser to an enclave it has verified; API answers name the attestation they ran under. (224)

## Topics

Artificial Intelligence, Privacy, Productivity, Developer Tools

## Links

- Website: https://undergrove.vercel.app
- Try without signing in: https://undergrove.vercel.app/#try (the small model in your browser) and https://undergrove.vercel.app/demo (the scripted tour)
- How it works: https://undergrove.vercel.app/architecture
- Side by side with other services: https://undergrove.vercel.app/compare

## Gallery order

1. `gallery/01-workspace.png`
2. `gifs/verified.gif`, or `gallery/02-what-the-relay-sees.png`
3. `gallery/03-model-race.png`, or `gifs/race.gif`
4. `gallery/04-python.png`, or `gifs/python.gif`
5. `gallery/05-architecture.png`
6. `gallery/06-compare.png`

Thumbnail: `thumbnail-240.png`. The footage in images 1 to 4 and in the GIFs is the real workspace with scripted answers, recorded against a stand-in model and a recorded attestation; each says so on the image. Keep those labels.

## Maker’s comment (draft)

Hi everyone. Undergrove is an AI workspace for open-weight models: GLM-5.3, Kimi K3, Gemma 4 31B, DeepSeek V4.1 Flash and GPT-OSS 120B.

What we care about most is where your words go. The models run in hardware enclaves operated by Tinfoil. Before the workspace sends anything, your browser checks the enclave’s attestation, the hardware’s signed statement of the code it’s running, then encrypts the request to a key held inside that enclave. That enclave runs Tinfoil’s router, which checks the model’s own enclave; your browser checks that one too where its hardware allows. Our relay sits in between for accounts and limits, and passes on ciphertext. Each answer can show its path and the exact bytes that left your tab.

What you can do with it:

- Web search with numbered citations, documents and codebases, images and voice
- Python on your spreadsheets, run in a worker in your browser tab
- Model races: one question, three models side by side, with speed and cost
- A canvas that shows revisions as tracked changes, and deep research reports
- Shared projects, end-to-end encrypted for your team
- An OpenAI-compatible API

Where it stands: live models are in a private alpha, free within a daily allowance for members. Anyone can try the workspace with sample answers, watch the scripted tour at /demo, or run a small model (LFM2.5 350M, a 324 MB download) in the browser on our home page with no sign-in. /architecture shows what each part of the system can and can’t see, with excerpts of the code, and /compare sets us beside ChatGPT, Claude, Mistral Vibe and Lumo using each one’s own pages.

What it doesn’t do yet: it hasn’t been independently audited. Our host, Vercel, receives ordinary request logs, and our relay sees each request’s size and timing. The developer API works differently from the workspace: our gateway reads API requests in memory to check them, and doesn’t store them. The enclave’s protections rest on the hardware and on Tinfoil’s published code.

We’d like your feedback on two things: whether the verification makes sense as shown, and which models you’d want next.

## Replies to likely questions (drafts)

**Is it open source?** Not yet. The architecture page shows excerpts of the code that handles your requests, with file paths and line numbers, and the enclaves run code that Tinfoil publishes as open source.

**What does it cost?** During the private alpha, members use live models free within a daily allowance. No payment is collected today. The pricing page has a calculator for the published per-model prices.

**Do you train on my conversations?** No. We don’t train models, and in the workspace we never hold conversation content we can read: conversations stay in your browser, and synced or shared copies are encrypted there first.

**Can I use it without an account?** Yes, with sample answers in the workspace, the small model in your browser on the home page, and the tour at /demo. Live models need alpha access.

**How is this different from a zero-retention promise?** A promise rests on the operator’s word. Here your browser checks the enclave’s attestation before sending, and the request is encrypted to that enclave, so our relay can’t read it. You still rely on the hardware and on Tinfoil’s published code, and what we can see is listed on the privacy page.