Provider Agreement
Draft, pending legal review. This is a good-faith draft written to reflect how Umbra actually operates during alpha; it has not yet been reviewed by counsel. Questions? Contact us.
This agreement is for people who run umbra host. As a provider you host public, platform-approved models on your own Mac, using your own Hugging Face key. You cannot read the prompts or outputs that flow through your machine, and the network is designed so that trust comes from attestation, not from a promise. This agreement applies to you in addition to the Terms of Service and Privacy Policy.
Experimental alpha with an HK-only payout rollout.Developer card top-ups are live. Provider cash-out is available only when the provider wallet says payouts are active and requires a payout-enabled Hong Kong Stripe connected account. The current Hong Kong platform cannot transfer funds to connected accounts in other countries. Only settled amounts shown as withdrawable may become payable; projections and promotional credit are not.
This is a draft, not legal advice. This document is a good-faith draft prepared for the operator and their counsel to review. It is not legal advice and does not create any attorney-client relationship.
1. Who this is for
This agreement covers anyone who enrolls a Mac as an Umbra provider and runs the provider client (umbra host). You must be at least 18 years old and able to enter into this agreement. A single Umbra account can act as both a developer (using API keys) and a provider (hosting models).
2. What you host, and with whose key
- You host public, open-weight models that the platform has approved for the catalog. You do not host private, secret, or buyer-supplied weights. Public model files are not treated as secret.
- You download model files using your own Hugging Face credentials. Umbra never receives your Hugging Face key, and the download happens under your account with Hugging Face.
- You are responsible for complying with each model's license, Hugging Face's terms, and the laws that apply to you. You confirm you have the right to download and serve each model you host.
3. Prompt privacy (you cannot read the work)
The core of Umbra is that you, the machine owner, cannot read the prompts or outputs that pass through your provider. The design assumes the machine owner may be adversarial and closes the software paths to the inference process rather than relying on your good behavior:
- Inference runs in-process (no inference server, socket, or IPC to intercept), and prompt and output buffers are zeroized after each request.
- The runtime is hardened against debuggers, memory reads, and binary tampering, and the network is designed to attest that this lockdown is in place (Secure Enclave, SIP, MDM/MDA, and code-identity).
- You agree not to attempt to read, capture, log, reconstruct, or exfiltrate prompts or outputs, and not to circumvent, disable, or tamper with the privacy or attestation mechanisms. Doing so is a material breach.
Accuracy note: production routing requires Apple hardware attestation and an app-targeted running-code identity challenge. Attestation verifies the provider client and machine state; it does not warrant the model's behavior or output.
4. Attestation and enrollment
Trust on the network is established by attestation, not by a promise. To host at the attested tier, you enroll your Mac (which can include MDM/MDA enrollment) so the operating system itself can report your device's security state. By enrolling you authorize the platform to verify your hardware identity and security posture on an ongoing basis. You agree to keep your machine in the attested configuration while hosting and not to spoof, replay, or falsify attestation signals.
5. Your responsibilities
- Run a genuine, unmodified provider client, and keep your machine, account, and credentials secure.
- Provide accurate device and account information (for example the machine name, chip, and unified-memory size used for routing and attestation).
- Be responsible for your own electricity, bandwidth, hardware wear, network connection, and local taxes or fees that may apply to you.
- Comply with the Terms of Service acceptable-use rules and all laws that apply to you.
You are primarily responsible for the models you choose to host. Because you selected, approved, and pulled each model with your own Hugging Face key, you are the party who put that model "into service" on the network. If a model you host produces dangerous, harmful, or illegal outputs, responsibility sits with you (the provider who chose to host it) and the model's author, not with Umbra, which is the neutral pipe that routes requests and verifies attestation. Umbra does not select, control, or endorse the models you choose to host; it surfaces each model's HF repo, license, pinned revision, and GGUF SHA-256 so developers can verify them independently. Attestation proves your machine is genuine; it does not make Umbra responsible for the model's behavior.
6. Earnings and Hong Kong payouts
The console may display base pay (for being attested, online, and ready) and on-demand earnings (per-token) so you can see how the model is intended to work. Payment eligibility is narrower than the figures displayed in the console:
- Only settled wallet earnings marked withdrawable are payable. Unsettled projections, estimates, promotional credit, and purchased developer credit cannot be withdrawn.
- Payouts require a completed, payout-enabled Hong Kong Stripe connected account, Stripe identity verification, any tax details Stripe requires, and a minimum withdrawable balance of $25.
- Umbra sends an approved withdrawal to your Stripe connected-account balance. Stripe controls the subsequent bank payout timing, supported bank accounts, verification, holds, and any Stripe fees.
- Displayed rates and projections may change prospectively during alpha. Your local taxes, reporting obligations, electricity, and other provider costs remain your responsibility.
7. Acceptable use for providers
The platform is a neutral marketplace. You independently select and host public, open-weight models from Hugging Face; the platform does not host, curate, or endorse specific models. You may choose to hostuncensored or unfiltered models —"uncensored" describes a category of model you MAY choose to host, not a feature of the platform. A model's presence in the catalog reflects your independent choice, not platform endorsement. You maynot host models that are clearly illegal, host models you lack the rights to serve, use the provider client to attack or de-anonymize users or the network, or use it to circumvent metering, routing, or access controls. The platform may remove a model from the approved catalog at any time.
8. You can stop anytime
Hosting is voluntary. You may stop hosting and unenroll your machine at any time, for any reason, with no penalty. Stopping does not convert promotional credit or projections into cash. Eligible settled earnings already shown as withdrawable remain subject to this agreement and Stripe verification. We may also suspend or terminate your participation at any time, including for suspected breach of this agreement, the Terms of Service, or applicable law.
9. As-is, no warranty
The provider client and the network are provided "as is" and "as available," without warranties of any kind, express or implied, including merchantability, fitness for a particular purpose, availability, accuracy, or non-infringement. The software is early-stage alpha and may change, break, or be discontinued. You run it at your own risk and are responsible for your own hardware.
10. Limitation of liability
To the maximum extent permitted by law, Umbra and its operators will not be liable for any indirect, incidental, special, consequential, or punitive damages, or for lost profits, lost data, hardware wear, or energy costs, arising from your operation of a provider machine. Our total liability is limited to the amount actually paid to you through Umbra in the prior twelve months.
11. Indemnification
You agree to indemnify and hold harmless Umbra and its operators from claims arising out of your operation of a provider machine, the models you choose to host, or your violation of this agreement, the Terms of Service, a model license, or applicable law.
12. Governing law
This agreement is governed by the laws of Hong Kong, without regard to conflict-of-laws rules. Disputes will be resolved in the courts of Hong Kong.
13. Changes
We may update this agreement as the network evolves. The "Last updated" date reflects the latest version; continued hosting after a change means you accept the updated agreement.
14. Contact
Questions about hosting? Reach us at [email protected].