Open source

AI runs the compute.
Local or remote.

Hypermesh is the open-source AI harness. People set the goals, the budgets and the limits. AI picks the model, finds the machine and runs the job, on hardware you own or on a host you lease.

It is also a privacy-first marketplace for local, edge, open-weight, neocloud and frontier models. Full Model is live on certified hosts. If no certified instance can serve a request, a listed public API can, and the receipt names that route. Choosing across those models, with each decision shown, comes in v1.

Become a host. Questions: hello@hyperme.sh. Hypermesh is for adults (18+).

Live. Pick a model from the catalog and lease an enrolled host. You pay in USDC on Base through Stripe, or buy USDC with a card through Stripe's onramp.
Coming in v1. Bring weights by URL or upload and run them by the hour on a certified class.

The harness

Conductor decides how a task runs. Visor is where the work happens. AIuditor decides what the task is allowed to do. You can also run the same system on your own hardware.

v0, in test

Conductor

Conductor is the harness and the optimization layer. It runs on your machine as a local proxy with OpenAI- and Anthropic-compatible endpoints. Cheap steps go to a cheap or local model. Hard steps go to a frontier model. Each request is checked with AIuditor.

  • A tool's normal model calls go directly to the provider with your own key. Only requests routed to a Hypermesh model use the gateway.
  • Coming in v1: model and host selection by profile, time budgets on leases, crash-recoverable runs, and bounded subagents. v1 starts with time budgets.

Conductor source

Open source

AIuditor

AIuditor is open source, and it is also available with an enterprise license with SOC 2 packaging. It applies your organization's policy and protects hosts.

  • Live: approved connections, an approval inbox, audit export, and a certification program. A private model upload reaches a lease only after a quarantined lab run, a sandbox report and a human verdict.
  • Coming in v1: a self-hosted install, a paid retention policy, and an org-wide gateway enforcement switch. On the public network, AIuditor stores an organization's requests only under that policy. Screening on its own does not store requests.

Linux desktop today

Visor

Visor is the open AI assistant for desktop and headless sessions. An agent can view, watch, listen, move the mouse and type, under rules you set.

  • Live on Linux desktop (x86_64 and ARM64), including a headless session on a machine with no monitor. Tool calls are approved or denied, then logged without arguments or secrets.
  • Coming in v1: more desktops, including macOS and Windows, headless builds for Alpine and Debian, and a phone app that can stand in for the phone assistant.

Linux downloads · Visor source

Templates live. Profile in v1

Deploy your own enterprise Hypermesh

Run Hypermesh on company hardware. Open-source deployment templates for Compose and Kubernetes are available now. The packaged enterprise profile comes in v1.

  • Coming in v1: internal token and hour accounting as the default for a self-hosted plane, and an accounting backend selector for pluggable billing. The profile also includes company hosts, quotas, usage exports and your identity provider.
  • Coming in v1: an AIuditor install you run yourself, a paid retention policy, and an org-wide gateway enforcement switch. Safety flagging and safety screening stay on the public network, not on a deployment you run.

Public source · Deployment templates · Architecture guide · Deployment guide

For help planning an internal deployment, write to hello@hyperme.sh.

Your prompts are yours

Hypermesh doesn't keep your prompts or outputs, train on them, or sell them unless you choose to.

  • Normal model calls from Conductor go directly to the provider with your own key. Only Hypermesh-routed calls use the gateway. Hypermesh does not hold your provider key.
  • Prompt research is opt-in. Personal information is removed before anything is used.
  • On the public network, AIuditor stores an organization's requests only under a paid retention policy. That policy comes in v1.
  • The public network screens for four categories under our acceptable use policy: child sexual abuse material, credible self-harm risk, malware or intrusion tooling aimed at others, and violent extremism. We report apparent child sexual abuse material to NCMEC as federal law requires. We don't monitor all traffic and can't guarantee that harmful content is caught.

Coming later, after v1

After v1

  • Physical machines. Cameras, sensors, mobile bases and arms, leased with the same checks and receipts as a model.
  • A network that organizes itself and heals. Hosts find each other, share load and route around a failure.

What you can rent

Pick a model from the Full Model catalog and check out. Sending requests is coming soon. Your Model, where you bring your own model, is coming soon too.

Full Model

Pick a model from the catalog and check out. It runs on enrolled hosts. Sending requests is coming soon.

Loading the list.

  • Nothing to upload. Pick a model from the catalog. Sending requests is coming soon.

Coming soon

Your Model (BYOM)

Bring your own model by URL or upload, and we'll check it and run it on the host you pick, by the hour. Not available yet.

Loading the list.

  • You'll pay by the hour for the host you pick.
  • You'll give us a model URL or upload, and we'll check it before it runs.
  • Not available yet.
Coming soon

Coming soon: run your own model on real hardware.

Hypermesh is building a way to run your own model on edge boards, local workstations and GPU hosts. You'll pick a compatible host, point to your model by URL or upload it, and we'll check it and run it there by the hour. That isn't available yet.

What you'll be able to do

  • Pick a compatible host from the list and see its hardware class and certified status.
  • Give us a link to your model, or upload it. We'll check it before it runs.
  • Get a receipt for the hours and the host.

What you can do today

  • Browse the hosts on /nodes.
  • Create an account and pick a model from the Full Model catalog. Sending requests is coming soon.
  • Have a system to test? Join host testing and help qualify your hardware for model hosting.

You rent compute for your model, not cameras, sensors, or a robot.

Create an account

Open models, on hardware you can name.

The Full Model catalog is open. Pick a model and check out. Sending requests and bringing your own model are coming soon.

Available now

  • Full Model. Pick a model from the catalog.
  • Speeds we've measured, nothing else. Where we haven't measured a model, the list says "We have not published a rate."

Coming soon

  • Sending requests. Coming soon.
  • Bring your own model. Pick a host and give us your model by URL or upload. We'll check it and run it on that host by the hour.

How it works

Create an account, pick a model from the catalog, and check out. You pay in USDC.

  1. Create an account.

    Create an account.

  2. Browse the hosts.

    See available hosts on /nodes. Each shows its hardware, whether it's certified, and whether it's online.

  3. Pick a model from the catalog.

    Choose a certified Full Model item and reserve it.

  4. Check out.

    You pay in USDC. No USDC yet? Buy it with a card or bank account through Stripe's crypto onramp.

  5. Tell us what broke.

    bugs@hyperme.sh

Coming soon: sending requests. Not available yet.

Coming soon: bring your own model. You'll pick a host, give us your model by URL or upload it, and we'll check it and run it on that host by the hour. Not available yet.

API access, with a key from your dashboard, is coming soon.

Host quickstart

Available now as test builds. The full steps are on How to enroll.

  1. Open the host dashboard.
  2. Pick Test and your processor.
  3. Run the one install command and say yes when it asks.
  4. Approve the box in your browser.
  5. Setup finishes on its own. Your box starts selling automatically once it's certified, and you can pause it from your dashboard. The first host of a new hardware class is reviewed by an admin first.

Edge boards, desktops and GPU rigs

Explore the hardware classes currently listed on Hypermesh. Our test program welcomes NVIDIA, AMD, Intel, Apple silicon and ARM systems, plus AI accelerator add-ons. Each system needs its own runtime and certification results before it can offer capacity.

Full Model

The live list

Posted prices. Pick a model and check out.

Loading the list.

Loading the list.

Have local AI hardware? Bring it to Hypermesh.

Start with a test build for your system, then help qualify it for hosting. We welcome edge boards, desktops, workstations and GPU rigs, including hardware our lab does not own. When live payouts open, you earn USD payouts to your bank through Stripe Connect for rented time your box delivers while it passes host checks. USDC payouts are after v0.

You earn on reserved hours

  • Renters will use it for Your Model when bring-your-model opens.
  • You are not paid up front. If your box drops out or fails host checks during a lease, you may not be paid for that lease.
  • The installer checks your hardware and asks before it changes anything. Your box can sell only after it passes certification, and you can pause selling at any time.

Host sign-up is open, and test builds of the host software are available now. Production builds are not released yet.

Help bring your hardware to Hypermesh

We welcome edge and local LLM hosts across NVIDIA, AMD, Intel, Apple silicon and ARM, including AI accelerators and M.2 modules. You can help test a system even when we do not own one.

Individual signed downloads in the test channel
PlatformHardware examplesCurrent test coverage
Linux ARM64Radxa ROCK 5, Orion O6N, Renesas V2H, JetsonDownload and inspect; setup and inference depend on the board and runtime. Vendor Yocto installation is still in development.
Linux AMD64 / x86-64AMD 395/495, NVIDIA and AMD GPU rigs, Intel systemsDownload and inspect; runtime support varies. Intel GPU/NPU adapters need further implementation and testing.
macOS ARM64Apple silicon Mac mini and StudioDownload and inspect. Native service setup is in development.

Accelerator support depends on the exact device, driver, runtime and model format. Coral and Radxa AICore modules are part of the testing roadmap; a matching connector does not guarantee that a model can run.

Hypermesh Terms

Accept these Terms before you download Hypermesh software or copy an install command. Terms of Service · Privacy Policy

Read the full Hypermesh Terms
# Hypermesh Terms

**Company:** Fyber Labs TN LLC, 116 Agnes Rd, Ste 200, Knoxville, TN 37919
**Contact:** legal@hyperme.sh · security@hyperme.sh · bugs@hyperme.sh · sales@hyperme.sh

By downloading, installing, or running this Hypermesh software, or by passing `--license accept` (or the matching environment accept), you agree to the Hypermesh Terms of Fyber Labs TN LLC. The sections below are part of those same Terms. Using the Hypermesh network (accounts, hosts, leases, payments, and related services) is under those same Terms. Host install consent for system changes (`--yes`) is separate and does not replace this acceptance.

---

## 4. Accounts and eligibility

- You must be at least 18 and able to form a binding contract. No one under 18 may use Hypermesh or create an account. If you use Hypermesh for a company, you confirm you're authorized to accept these terms for it, and "you" includes that company.
- You can sign up with an email address and password, or with a Microsoft, GitHub, or Google account. Give accurate information and keep it up to date.
- You're responsible for everything done through your account, including by the CLI, Visor, or any tool you connect with your sign-in. Keep your credentials safe and tell security@hyperme.sh right away if you think your account has been compromised.
- You may not use Hypermesh if you're barred from doing so under US law, including US sanctions (section 14).

---

## 11. Our software

We give you a limited, personal, non-exclusive, non-transferable, revocable license to install and use the host software, the CLI, and Visor to use Hypermesh under these terms. Where we publish a component under an open-source license, that license controls that component. The host software may update itself or ask you to update. Software downloads may come from GitHub or our own storage, and you should check them against our published checksums.

---

## 16. Disclaimers

**THE SERVICE, HOST SOFTWARE, CLI, VISOR, MODELS, OUTPUTS, AND ALL CONTENT ARE PROVIDED "AS IS" AND "AS AVAILABLE". TO THE MAXIMUM EXTENT ALLOWED BY LAW, FYBER LABS DISCLAIMS ALL WARRANTIES, EXPRESS OR IMPLIED, INCLUDING MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE, TITLE, AND NON-INFRINGEMENT.** We don't promise the Service will be uninterrupted, secure, or error-free, that any host will be available or perform at any speed, that certification means anything beyond passing our checks at the time, that outputs will be accurate, or that installing the host software won't affect your hardware or other software. We don't publish uptime or performance commitments. You accept the risk of ordinary errors, downtime, and omissions.

---

## 17. Limitation of liability

Our liability works in two layers: one for the software and services we run, and one for host hardware.

**Our software and services.** This covers the portal, the CLI, Visor, the host software, and the other services Fyber Labs runs. **TO THE MAXIMUM EXTENT ALLOWED BY LAW, OUR TOTAL LIABILITY FOR ALL CLAIMS RELATING TO THEM IS LIMITED TO THE GREATER OF (1) THE FEES YOU PAID US IN THE 12 MONTHS BEFORE THE CLAIM AROSE OR (2) US$100.**

**Host hardware.** For anything about a host's hardware, Fyber Labs acts only as the marketplace that matches renters and hosts. Hosts own and run the hardware, not us. **TO THE MAXIMUM EXTENT ALLOWED BY LAW, FYBER LABS ISN'T LIABLE FOR DAMAGE TO OR FAILURE OF HOST HARDWARE, FOR HOSTS THAT ARE OFFLINE OR SLOW, FOR POOR NETWORK PERFORMANCE, OR FOR DATA EXPOSED ON A HOST'S MACHINE. FYBER LABS ALSO ISN'T LIABLE FOR PERFORMANCE CHANGES AFTER OUR CHECKS RAN, INCLUDING THROTTLING, THERMAL LIMITS, NETWORK LATENCY, POWER OR LOCAL OUTAGES, AND HARDWARE WEAR OR DEGRADATION, OR FOR A MODEL THAT FAILS TO LOAD, RUNS SLOWLY, OR PRODUCES CORRUPTED OR WRONG OUTPUT BECAUSE OF A HOST'S HARDWARE. CERTIFICATION AND BENCHMARK RESULTS DON'T CHANGE THIS. IF A LEASE FAILS, ISN'T DELIVERED, OR UNDER-DELIVERS, FOR ANY REASON, YOUR SOLE AND EXCLUSIVE REMEDY IS THE REFUND, IF ANY, DESCRIBED IN SECTION 8.**

**Digital assets.** **TO THE MAXIMUM EXTENT ALLOWED BY LAW, FYBER LABS ISN'T LIABLE FOR LOSSES CAUSED BY CHANGES IN THE PRICE OF USDC OR ITS PEG TO THE US DOLLAR; NETWORK FEES OR SPIKES IN GAS FEES; BLOCKCHAIN NETWORK CONGESTION, FORKS, OR FAILURES; DELAYS IN SENDING OR RECEIVING PAYMENTS, PAYOUTS, OR REFUNDS CAUSED BY THE BLOCKCHAIN; FAILURES BY THE STABLECOIN ISSUER OR THE ONRAMP PROVIDER; MISTAKES IN A WALLET OR ADDRESS, INCLUDING SENDING TO OR REGISTERING THE WRONG ADDRESS; OR CHANGES IN LAW OR REGULATION THAT AFFECT DIGITAL ASSETS.**

**In all cases, to the maximum extent allowed by law:**
- **FYBER LABS WON'T BE LIABLE FOR ANY INDIRECT, INCIDENTAL, SPECIAL, CONSEQUENTIAL, EXEMPLARY, OR PUNITIVE DAMAGES, INCLUDING LOST PROFITS, REVENUE, OR GOODWILL, LOST OR CORRUPTED DATA, AND BUSINESS INTERRUPTION, EVEN IF WE WERE TOLD THEY WERE POSSIBLE.**
- **OTHER LIMITS IN THESE TERMS, INCLUDING FOR VISOR, DON'T LOWER THE SOFTWARE-AND-SERVICES CAP BELOW US$100. THE ONLY EXCEPTION IS A FAILED OR UNDER-DELIVERED LEASE, WHERE THE REFUND IS YOUR ONLY REMEDY.**

Some places don't allow these limits, so they may not fully apply to you.

---

## 21. Governing law and courts

These terms are governed by the laws of the State of Tennessee, without regard to conflict-of-law rules, except that the Federal Arbitration Act governs section 22. Any claim that isn't arbitrated under section 22 (other than a small claims case) will be brought only in the state courts located in Knox County, Tennessee, or the United States District Court for the Eastern District of Tennessee. Both parties consent to personal jurisdiction in those courts. This includes claims brought in court under the intellectual property and injunctive relief carve-out, claims severed under the severability rule, and claims by anyone who opted out of arbitration.

---

## 22. Dispute resolution, binding arbitration, and class-action waiver

**Please read this section carefully. It affects your legal rights, including your right to go to court.**

**(a) Informal resolution first.** Before starting arbitration or a court case, the party with a dispute must send the other a written notice describing it and the relief wanted. Send notices to us at legal@hyperme.sh or by mail to Fyber Labs TN LLC, 116 Agnes Rd, Ste 200, Knoxville, TN 37919. We'll send notices to you at the email address we have for you. Both of us will try in good faith to resolve the dispute for 30 days after the notice is received. Deadlines to bring a claim are paused during those 30 days.

**(b) Binding individual arbitration.** If the dispute isn't resolved within those 30 days, you and Fyber Labs agree that any dispute, claim, or controversy arising out of or relating to these terms or Hypermesh (including the Service, the host software, the CLI, and Visor) (a "Dispute") will be resolved by final, binding arbitration on an individual basis, except as set out in (d) and (e).
- The American Arbitration Association (AAA) will administer the arbitration.
- If you're an individual using Hypermesh (including the Service, the host software, the CLI, and Visor) for personal, family, or household purposes, the AAA Consumer Arbitration Rules apply. If you're a business, the AAA Commercial Arbitration Rules apply. The rules are available at adr.org.
- A single arbitrator will decide the Dispute. The arbitrator can award the same individual relief a court could, but only in favor of the individual party bringing the claim and only to the extent needed for that party's individual claim.
- The arbitrator decides questions about the scope, validity, and enforceability of this arbitration agreement, except that a court decides questions about the class-action waiver in (f).

**(c) Where it happens and fees.** Any in-person hearing will be held in Knox County, Tennessee. If you're a consumer, you can instead choose to have the arbitration conducted by video or telephone, or decided on written submissions. Filing, administration, and arbitrator fees are paid as the applicable AAA rules provide.

**(d) Small claims.** Either party may bring an individual claim in small claims court instead, if the claim qualifies and stays in that court on an individual basis.

**(e) Intellectual property and injunctive relief.** Either party may go to court to address infringement or misuse of intellectual property rights, or to seek temporary or preliminary injunctive relief to prevent irreparable harm while an arbitration is pending.

**(f) Class-action and jury-trial waiver.** You and Fyber Labs each agree to bring Disputes only in an individual capacity, and not as a plaintiff or class member in any class, collective, consolidated, or representative proceeding. The arbitrator may not consolidate more than one person's claims or preside over any class or representative proceeding. **YOU AND FYBER LABS EACH WAIVE THE RIGHT TO A JURY TRIAL.**

**(g) 30-day opt-out.** You can opt out of this arbitration agreement within 30 days after you first accept these terms (for example, by creating an account, enrolling a host, checking out, or downloading our software). To opt out, email legal@hyperme.sh with your name, the email address you use with us, and a clear statement that you want to opt out of arbitration. Opting out doesn't affect any other part of these terms. If you opt out, section 21 applies to Disputes.

**(h) Severability.** If the class-action waiver in (f) is found unenforceable for a particular claim or request for relief, that claim or relief, and only that, will be severed from the arbitration. It will be brought in the state courts located in Knox County, Tennessee, or the United States District Court for the Eastern District of Tennessee. If any other part of this section is found unenforceable, the rest of it still applies.

**(i) Federal Arbitration Act.** These terms involve interstate commerce. The Federal Arbitration Act (9 U.S.C. § 1 et seq.) governs the interpretation and enforcement of this section, notwithstanding the Tennessee governing-law clause.

**(j) Changes to this section.** If we change this section, the change won't apply to a Dispute that either party gave notice of before the change took effect.

To accept without a prompt, run hypermesh-agent enroll --license accept, or set HYPERMESH_LICENSE_ACCEPT=yes in agent.env. Host install --yes does not accept these Terms.

Accept the Hypermesh Terms to download a test build.

Download, inspect and report what works or fails. Help us progress from bring-up to Bronze certification and end-to-end model usage as those services become available. When live payouts open, you earn USD payouts to your bank through Stripe Connect for rented time your box delivers while it passes host checks. USDC payouts are after v0.

Looking ahead: shared memory across hosts

Our shared-memory roadmap aims to let models use memory across multiple machines through a shared address space. Ethernet, NVIDIA systems and Apple Thunderbolt are in scope. This capability is planned and is not available yet.

Open software. Open market. Shared compute. The software is free, and the network runs on hosts and models that people choose to share.

Open source, end to end. Shared all the way down.

All of the Hypermesh software is free and open source under Apache 2.0 and MIT, at github.com/hyperme-sh. That includes the host agent, the gateway, the control plane, the CLI, the auditor and open payment options you can plug in or swap out. You can read the code, run it yourself or fork it.

Everyone on the network shares something:

  • We share the software. The harness, the routing and the tools are free for anyone to run, extend or self-host.
  • We share the market. Any model, host or agent can join the Hypermesh network on the same terms.
  • Hosts share their hardware. Edge boxes, workstations and racks earn by serving models when they'd otherwise sit idle.
  • Builders share their models. Open, fine-tuned and frontier models are listed side by side, so agents can pick the right one for each job.

Pricing

Each hardware class and model gets a posted price. You pay in USDC, or buy USDC with a card through Stripe. When live payouts open, hosts earn USD payouts to their bank through Stripe Connect for rented time their box delivers while it passes host checks. USDC payouts are after v0.

Loading the list.

See pricing