A routable AI compute fabric,
not another single-cluster tool.
mlxMesh federates geographically-spread machines into one distributed inference network — with strict privacy tiers, measured (not declared) performance accounting, and no native token.
Why this exists
The infrastructure gap
The industry's answer to AI demand is more concentration: hundreds of billions in debt-financed data centers, sized for real-time frontier-model queries — a workload pattern that's a minority of how AI actually gets used in production. Most real usage is background and asynchronous: a cron job reading server metrics, a batch summarization pass, an anomaly check. None of it needs a $44M/MW data center.
Meanwhile over 40% of the world's GPU capacity sits idle at any given moment — and millions of Apple Silicon Macs, with genuinely fast unified memory, go unused overnight. mlxMesh exists to connect those two facts.
The trust problem
Tools like Exo turn one LAN into a single inference cluster. Federating clusters across the open internet isn't a bigger version of that problem — it's a trust problem: routing jobs to machines you don't operate, verifying they actually did the work, and paying out fairly, with no shared owner in the middle.
The usual answer is a blockchain and a token — let market speculation fund the security budget. That pattern tends to decouple token price from actual network utility, then collapse when sentiment turns. mlxMesh borrows the idea decentralized ledgers got right — verify cryptographically, don't trust an operator's word — without adopting a speculative asset to do it. Credits are internal ledger accounting, settled off-chain, sized to compute actually delivered.
What it is
Most distributed inference tools (e.g. Exo) work within a single LAN cluster. mlxMesh adds a coordination layer above that: it federates multiple clusters across the internet into a routable mesh.
Dual-lane routing
Fast lane for interactive jobs (resolver-routed, low-latency), background lane for recurring/batch jobs (scheduler-routed, sticky-session).
Division-order accounting
Measured resource lines, not declared promises — credits from bootstrap grants decay as earned capacity grows.
Sensitivity tiers
LOW / MODERATE / HIGH_REQUIRES_ATTESTATION, with a Secure Enclave gate on Apple Silicon for the strictest tier.
Ed25519 node identity
Every node's identity is derived from its public key — never operator-chosen, never spoofable.
iOS coordination layer
iPhone/iPad devices classify on-device and host encrypted payload pointers, adding a privacy layer without becoming compute nodes.
Portable wallet identity
An Ed25519 account key, iCloud-Keychain synced and seed-recoverable, that consolidates credits across a user's devices.
Native SwiftUI clients for iOS/iPadOS, tvOS, and watchOS render live topology and drive contribution/coordination directly from Apple hardware.
How it works
Under the hood, mlxMesh leans on a handful of ideas borrowed from fields that solved these exact problems long before distributed AI existed.
Measured, not declared
Oil & gas division orders allocate revenue by measured production at the wellhead, not a well's theoretical output. mlxMesh settles the same way — a node's payout tracks independently verified tokens delivered on a given job, not whatever throughput it claimed at registration.
A cryptographic chain of custody
Every node's identity is its own public key — nothing to spoof, nothing an operator picks. Privacy-sensitive job payloads are sealed end-to-end before they ever reach routing infrastructure, and coordinator-to-node dispatch pins the exact certificate seen at registration — the same trust-on-first-use model SSH uses, since a permissionless network has no shared certificate authority to hand out.
Backpressure, not brute force
Every pod tracks queue depth and in-flight jobs the way message brokers like MQTT signal saturation back through a pipeline — instead of silently dropping jobs or piling up an unbounded queue, the network throttles and reports backpressure so it degrades predictably under load instead of falling over.
Sybil resistance without a paywall
Free bootstrap credits are gated behind a small proof-of-work challenge — the same shape of idea Hashcash and Bitcoin popularized — so minting disposable identities to farm credits costs real computation, not just a button click.
Open source, AGPL-3.0
mlxMesh is licensed under the GNU Affero General Public License v3.0. Free to use, modify, and distribute under AGPL-3.0 — network users (SaaS) get access to the source, and derivative works stay AGPL-3.0. Commercial use outside those terms (proprietary SaaS, closed-source integration) requires a separate commercial license.