// //
OWNED BY NO ONE.
Uncensored AI inference, end-to-end encrypted messaging, and formally verified mathematics — served on GPUs that people contribute, not corporate data centers. When the network proves something, it proves it in Lean 4 via Harmonic's Aristotle, then signs the result so you can check it yourself.
UNCENSORED INFERENCE
OpenAI-compatible API with no refusal layer and no prompt logging. Your prompts aren't stored — only billing is.
PRIVATE BY CONSTRUCTION
Direct messages are sealed on your device with X25519 + XChaCha20-Poly1305. The relay only ever handles ciphertext.
PROVABLE, NOT PLAUSIBLE
Ask for a theorem and you get a Lean 4 proof checked by machine — then signed with ed25519 so anyone can verify it came from us.
DECENTRALIZED & REWARDED
Contribute a GPU from a browser tab or a native worker and earn 70% of the credits your work serves.
Most AI answers are plausible.
These are proved.
A language model that sounds certain is still guessing. So for mathematics we don't ask a model to be convincing — we route the problem to Harmonic's Aristotle, which writes the argument in Lean 4, a proof assistant that refuses to compile anything that doesn't follow. If it builds clean, with no sorry, no admit and no new axioms, the result is a fact, not an opinion.
Then we go one step further: every such run is stamped with a run marker and signed with ed25519. The published proof page verifies that signature in your browser — you never have to take our word for it.
⚠ one documented exception · Aristotle is an external API, so selecting that model sends the problem to a third party. Every such step is labelled leaves network in the interface. Everything else runs inside the network.
Relay privacy, machine-checked
Not a claim — a receipt. We asked the network to formalise our own privacy property and prove it. It was checked by machine, not asserted by a chatbot, and the whole trail is public: every command, every file edit, the build, and the final check for sorry or admit.
- ✓any two plaintexts produce identical relay views for fixed visible parameters
- ✓every deterministic observer using only that view returns the same result
- ✓for distinct messages, no view-only reconstruction function can be correct for both
assumes · an explicit FreshSealPlaintextIndependent idealisation, stated as a hypothesis — not claimed as a property of ordinary deterministic AEAD. It proves the model, not the running code — and the verification was performed by Harmonic's Aristotle, orchestrated and attested by us.
Reserved before a job dispatches and settled on completion, inside one atomic transaction. No double-spend, no negative balances.
Whoever serves the tokens keeps the majority of what the work is worth. Contribute from a browser tab or a native worker.
Every account gets its own Ethereum deposit address, so a payment is bound to you by where it lands. Priced by the Pyth oracle, verified on-chain, and credited once per transaction.
Gates holder-only features like the X bot's /prove, and funds buyback-and-burn. A published design — not yet minted or tradeable.
BEYOND A COMPUTE NETWORK
Other decentralized-inference projects stop at chat. Open Communication puts private human messaging on the same rails — one identity, one connection, one economy — and adds something none of them have: answers you can check, signed and machine-verified.