About Tao Conclave

A private AI chat app with three clear modes: Private (confidential TEE), Encrypted (browser-verified E2EE), and Auto (assistant mode that may use web search automatically).

What it does today

Tao Conclave is built for useful AI without pretending every mode proves the same thing. The app does not store your chat on the server.

• Private — messages are relayed to Chutes models running in verified secure hardware (confidential TEE). Conclave can transiently see plaintext in this mode, so the claim depends on HTTPS, no logging, and hardware isolation. Optional web research asks you to confirm the exact Brave query before it leaves.
• Encrypted — your browser encrypts to a key bound to attested hardware. Conclave relays ciphertext only, and each encrypted turn can produce a receipt. Web search is off in this mode.
• Auto — answers directly on Chutes confidential TEE when it can, or sends a focused query to Brave automatically when web search is useful. No confirmation step.
• Local history — if you turn on saved history, it stays in this browser only.
• Model choice — the picker shows models that are eligible for the current mode.

Don't trust. Verify.

For Encrypted turns, the receipt checker recomputes the key binding from raw TDX quote data, checks the ciphertext hash, and can be verified outside Conclave with the Chutes-documented Intel/NVIDIA attestation path.

See Don't Trust. Verify. for the plain-language proof model.

Web search boundaries

Any turn that searches the web crosses one boundary: your query is sent to Brave Search to retrieve sources. Inference and synthesis still run on Chutes TEE unless you are in Encrypted mode (where web search is unavailable).

• Private + web research — the private model may propose a search; you see and can edit the exact query before anything goes to Brave.
• Auto — the agent decides: answer directly, or search Brave automatically when fresh sources help. No confirmation step.

Conclave shows which path each answer actually used. Free web search is limited to protect the shared search budget; private chat remains open.

What is not included yet

• Fully independent frontend provenance — frontend hashes are published, and a public source mirror is prepared. Do not treat this as full independence until the public repo is live and hash-matched.
• Multi-model rooms and consensus — planned as an advanced surface, not the current MVP.
• Server-side memory — not implemented.

Operator telemetry

We run unlisted benchmark telemetry to watch model latency, reliability, web search readiness, and cost. It uses fixed test prompts only and never contains user chat.

See Privacy for the exact privacy boundaries.