██╗ ██╗███████╗██████╗ ██████╗ ██████╗██╗ ██╗ █████╗ ████████╗
██║ ██║██╔════╝██╔══██╗██╔═══██╗ ██╔════╝██║ ██║██╔══██╗╚══██╔══╝
██║ ██║█████╗ ██████╔╝██║ ██║ ██║ ███████║███████║ ██║
╚██╗ ██╔╝██╔══╝ ██╔══██╗██║ ██║ ██║ ██╔══██║██╔══██║ ██║
╚████╔╝ ███████╗██║ ██║╚██████╔╝ ╚██████╗██║ ██║██║ ██║ ██║
╚═══╝ ╚══════╝╚═╝ ╚═╝ ╚═════╝ ╚═════╝╚═╝ ╚═╝╚═╝ ╚═╝ ╚═╝
v e r o · c h a t 🔒
A self-hosted, private ChatGPT-style app where the language model runs inside a hardware-encrypted Trusted Execution Environment (TEE). Sign in with Google, chat with a model whose weights and your prompts are encrypted in use, and verify — down to a container digest — exactly what code is running.
Start a chat, upload a confidential contract, and ask about its terms — then click Verify hardware and watch the browser check an Intel-signed TDX attestation quote on the spot:
The clip was captured on the auth-free
demo-frontend-noauthbranch (a frontend-only build with authentication removed and the chat/attestation APIs mocked, so the UI can be recorded without Google OAuth or the TEE model backend). The attestation panel mirrors the real TDX output byte-for-byte.
This is the headline feature. The app is built in the open by GitHub Actions and published to GHCR. That public build produces an image digest, and the very same digest is what the running app shows in its footer — with each answer tagged by the git commit that generated it.
🔒 Currently deployed in the TEE
image ghcr.io/505labs/confidential-chat@sha256:f5eb3c4f17e10b82c616a9f22e55a042b6f8e99bb83bb0062f2b1cdfbe6f5baf
commit 1de27ffcdf3c3f73c8ffa16c98d4c63ea906e547
built 2026-07-29T12:26:37Z (UTC, by GitHub Actions)
Pull this exact image: docker pull ghcr.io/505labs/confidential-chat@sha256:f5eb3c4f17e10b82c616a9f22e55a042b6f8e99bb83bb0062f2b1cdfbe6f5baf — the digest above is what the running app shows in its footer.
The chain of custody:
public source ─▶ public GitHub Actions build ─▶ ghcr.io/…@sha256:DIGEST
│ │
│ ▼
└────────── same DIGEST shown in the app footer ◀── deployed into the TEE
each reply tagged with its commit SHA
You don't have to trust a screenshot — pull the exact image yourself:
docker pull ghcr.io/505labs/confidential-chat@sha256:<digest>.
- 🔐 Model runs in a TEE — Intel TDX encrypts the VM's RAM in hardware (with integrity protection); prompts and weights are protected in use, not just at rest.
- 🧾 Provable hardware — the app produces + self-verifies an Intel TDX attestation quote against Intel's root of trust (TDREPORT → TD Quote → DCAP), with no cloud provider in the verification chain.
- 🪪 Google sign-in — real per-user accounts via Google OAuth. Anyone with a Google account can sign in and use the app; API routes are per-IP rate-limited.
- 💾 Local chat history — a lightweight SQLite DB on the VM. No external database, no data leaving the box.
- 🧬 Verifiable builds — public CI → image digest → shown in-app + in this README; each reply carries the code commit hash.
- 🌐 Real HTTPS, no domain — automatic Let's Encrypt certs via
sslip.io.
GCP Confidential VM (Intel TDX Trust Domain · RAM encrypted + integrity)
Public user ┌──────────────────────────────────────────────────────┐
│ HTTPS :443 │ caddy ── reverse proxy, auto Let's Encrypt │
└───────────────────┼─▶ app ── Next.js chat UI │
│ ├─ Google OAuth (Auth.js) │
│ ├─ SQLite (accounts + conversation history) │
│ ├─ attestor ── TDX quote via configfs-tsm, │
│ │ self-hosted DCAP (dcap-qvl) │
│ └─▶ llama.cpp ── CPU inference │
│ └─▶ Qwen2.5-1.5B (GGUF) │
└──────────────────────────────────────────────────────┘
Only :80/:443 exposed. Accounts + conversations never leave the VM.
Footer shows the deployed image digest; each reply shows its code commit.
Full design + decision log: docs/ARCHITECTURE.md.
# 1. Provision the Intel TDX Confidential VM (static IP + firewall)
gcloud config set project YOUR_PROJECT_ID
MACHINE=n2d-highcpu-8 ./infra/create-vm.sh # prints the VM IP + PUBLIC_HOST
# 2. Create a Google OAuth client (Web) in the Cloud Console.
# Authorized redirect URI: https://<PUBLIC_HOST>/api/auth/callback/google
# 3. Deploy the app (pulls the CI-built image by digest, wires up Caddy + llama)
PUBLIC_HOST=<ip-with-dashes>.sslip.io ./deploy/deploy-app.shThen open https://<PUBLIC_HOST> — any Google account can sign in and start chatting.
Verify the TEE is genuine anytime:
./infra/verify-attestation.sh # pulls a TDX quote, verifies it to Intel's root with dcap-qvl| Path | Purpose |
|---|---|
app/ |
The custom Next.js chat app (Auth.js + SQLite + streaming). |
app/Dockerfile |
Multi-stage build → the image published by CI. |
.github/workflows/build.yml |
Public build → push to GHCR → update this README's digest. |
scripts/update-readme-digest.sh |
Rewrites the digest block above. |
deploy/docker-compose.vm.yml |
VM stack: llama.cpp + app (pinned by digest) + Caddy. |
deploy/deploy-app.sh |
Resolve the GHCR digest, template compose, deploy to the VM. |
infra/create-vm.sh |
Provision the Confidential VM + static IP + firewall. |
infra/verify-attestation.sh |
Request a TDX quote + verify it to Intel's root (self-hosted DCAP). |
docs/ARCHITECTURE.md · docs/DEPLOY.md |
Design/decision log · deploy runbook. |
The app reads its config from environment variables (see app/.env.example):
| Variable | What it does |
|---|---|
AUTH_SECRET |
Auth.js session secret (openssl rand -hex 32). |
AUTH_GOOGLE_ID / AUTH_GOOGLE_SECRET |
Google OAuth client credentials. |
LLAMA_BASE_URL / LLAMA_API_KEY |
llama.cpp endpoint + key (inside the TEE). |
DB_PATH |
SQLite file path (mounted volume, defaults to /data/app.db). |
IMAGE_DIGEST |
Injected at deploy so the footer shows the running image's digest. |
Secrets never live in the repo — only in the deploy environment. Regenerate them per deployment.
For a larger model, run on a confidential-mode NVIDIA GPU (H100/H200) alongside the TDX CPU Trust Domain — the GPU produces its own NVIDIA-signed attestation that composes with the TDX quote. The CPU attestation flow is unchanged; the demo here is CPU-only to keep costs down.
MIT © 2026 Snojj25