Draft: not yet legally reviewed
Terms of service
Draft last updated 20 October 2026
1. The service
Amber is a hosted service that runs Model Configurations on Tasks in sandboxed Environments, on public and private Benchmarks, and reports Rewards and costs. These terms apply when you use the Amber website, CLI or API.
2. Your Workspace
You sign in with GitHub or Google. Signing in creates a personal Workspace. You are responsible for what happens in your Workspace, including the actions of its members and anything done with its API keys. Keep API keys secret; they are shown only once.
3. Wallet and fees
- Jobs are paid from your Wallet, a prepaid balance. Amber charges the pass-through cost of Models and sandboxes plus a platform fee, currently 12.5%.
- Before a Job starts you are shown a Quote. The Quote is reserved from your Wallet and is the most the Job can spend, though in-flight Trials may overshoot it and Amber absorbs that overshoot.
- A new Workspace starts with no credit. Until paid top-ups launch, we add credit to a Wallet on request, at our discretion. Credit we add has no cash value, and we may reverse credit we believe was obtained by abuse.
- Trials that remain Errored after retries are not charged.
- Paid top-ups are not available yet. Terms for them will be added before they launch.
4. Your content
You keep ownership of the Tasks, files and other material you upload, including private Benchmarks. You give Amber permission to store and run them, and to send them to the Providers that run Models for your Jobs, solely to provide the service to your Workspace. Only upload material you have the right to use for this.
Uploaded Tasks are untrusted code. They run only in isolated sandboxes, never on Amber’s own servers.
5. Acceptable use
- Do not attack, probe or disrupt the service or other users.
- Do not use Tasks or Agents to attempt to escape a sandbox, reach networks you are not permitted to, or abuse the model proxy.
- Follow the terms of the Providers whose Models you run through Amber.
6. Results
Rewards and comparisons are statistical measurements on specific Benchmark Versions and Agent Versions. They do not guarantee how a Model Configuration will behave on other work. Confidence intervals and “too close to call” flags are part of the result; please read them.
7. Availability and changes
Amber is early-access software provided as is, without warranties. We may change, limit or stop features, and may suspend a Workspace that breaks these terms. Transcripts are deleted after 90 days.
8. Liability
To the extent the law allows, Amber is not liable for indirect or consequential losses, and its total liability is limited to the amount you paid for the service in the 12 months before the claim. This section needs legal review before it is final.
9. Contact
Questions about these terms: open an issue. A dedicated contact address will be added before launch.