Distributed LLM inference. Run the client to use the network, enable peer mode to serve it, or run the agent headless and only earn.
A distributed inference network with one application and two roles. The desktop client profiles the hardware on first launch, reports which models it can serve at acceptable latency and the approximate credit rate, then places a live counter in the system tray. Credits move in both directions on a single ledger — purchased, earned, spent and cashed out are kept as separate line items, so a machine that both serves and consumes settles against one balance while the two sources stay distinguishable. Inference covers quantised 3B to 14B models; training is out of scope for this generation. Unlike the token-based networks in the same space, settlement is a credit ledger rather than a cryptocurrency, which keeps the regulatory surface small and the accounting ordinary.
Chat only through the browser with no installation; peer and chat together through the desktop client; or peer only as a background service under systemd, launchd or a Windows service. The peer agent is a single Go binary and ships with a Helm chart for cluster deployment.
First launch profiles the GPU, reports which models it can serve within acceptable latency and estimates the hourly credit rate. A tray counter shows served requests and accumulated credits without opening the application.
Purchased, earned, spent and cashed-out credits are separate entries against one account. Operators who both serve and consume see a net position, while earned and purchased balances remain separable for tax treatment and future eligibility criteria.
Channel messages use XChaCha20-Poly1305 under a per-channel symmetric key, sealed with libsodium sealed_box to each member's X25519 public key, which is itself derived from the account password through Argon2id. The server holds ciphertext and sealed blobs only; keys never leave the operator's devices.
The relay removes the user identifier, IP address and conversation identifier before a request reaches a peer. Repeated requests are distributed across peers wherever capacity allows, so no single peer can reconstruct a usage profile.
The network-wide catalogue is curated centrally, but a self-operated peer decides what it serves. A privately registered model is visible only to the registering account unless it is explicitly advertised to the network.
70% of every token is retained by the peer, 25% funds the platform and 5% goes to a refund and risk pool. Payouts run through Wise once the KYC threshold is met — no exchange listing, no wallet and no token price to track.
The encryption is not bolted on to a chat product — it is the product. Three concrete promises, and three things it would be lying about if it claimed more.
Channel messages are encrypted with XChaCha20-Poly1305 under a per-channel symmetric key. That key never reaches the server in plaintext: it is sealed with libsodium's sealed_box to each member's X25519 public key, which is itself derived from your password via Argon2id. The server holds ciphertext and sealed blobs. The keys live on your devices.
Channels have zero server-side persistence — messages pass through the signalling bus as ciphertext and are never written to disk. Direct messages are the one exception: an opaque queue holds them for up to seven days so an offline recipient does not lose the thread, then it is deleted automatically.
When a prompt goes to a peer's GPU, the relay strips your user id, IP and conversation id first. The peer sees the prompt — it has to, to answer — but cannot link it back to you. Repeated requests rotate across peers where possible, to make profile reconstruction harder.
| CoralMesh | Discord / Slack | Telegram | Signal | |
|---|---|---|---|---|
| Server can read messages | No | Yes | Yes | No |
| Messages stored long-term | No — DMs 7d, channels 0 | Forever | Forever | No |
| AI bot inside channels | Built in | 3rd-party only | 3rd-party only | — |
| GPU earnings for users | Yes | — | — | — |
An estimate, not a promise. Throughput is averaged per GPU class on a ~7B-class model, and the electricity comes straight back off the top.
≈40 tok/s · ≈200 W while serving
Hours the machine is on and you are not using it.
Your rate. Drag it to what you actually pay.
per month, after electricity
Real earnings depend on demand for the models you serve. 1 credit = $0.01. Peers keep 70%, the platform takes 25%, and 5% goes to a refund and risk pool. Cashouts above the KYC threshold require ID verification.
Spare H100 or RTX racks during off-peak hours? Mark 5–20% of the fleet as available and CoralMesh routes inference into it, in 30-second slots, with no sales motion and no customer onboarding of your own. Data-centre-grade peers rank first in the scheduler and earn a 1.2× quality multiplier. The agent is a single Go binary with a Helm chart; resource caps and thermal hooks are built in, so it will not fight your existing scheduler. Fleets of 50+ GPUs can negotiate committed minimums, custom SLAs and a dedicated routing pool.
enterprise@coralmesh.io →Any Ollama-compatible model. The network-blessed catalog is curated, but if you run your own peer you decide what it serves.
No. Chats are routed to peers. You only need a GPU if you want to earn credits by serving the network.
It depends on the card. A modern 200 W GPU running 8 hours a day at $0.15/kWh costs about $7 a month — the calculator above subtracts exactly that.
Sharing idle compute is legal in the EU, US, UK and Türkiye. Some jurisdictions tax peer earnings as income, so check locally. Cashouts above the KYC threshold require ID verification.
The gateway retries transparently on another peer running the same model. The user sees a brief stall, never a half-finished response.
Yes. Register it on your peer and only your account sees it in the menu, unless you choose to advertise it to the network.
Earn credits, reach the $500 KYC threshold, submit ID, and operations processes a Wise transfer to your registered email within 14 days during the beta.
Portfolio products serving the same audience.
No per-seat licensing, no vendor lock-in, no unpredictable invoicing. Our teams are available to scope an evaluation for any product in the portfolio.