HAL SUPREME
OpenAI-compatible · bounded by design

Bring your tools.
Keep your keys.

Point an OpenAI-compatible client at HAL, inspect anonymous health, or run the gateway on your own machine. Protected model, chat, and peer routes require an operator-issued bearer token; this page never asks you to paste one into the browser.

One surface, explicit doors.

Discovery is not permission. The API reports health openly, while protected capability routes require authentication and keep provider credentials server-side.

Public

GET /health

Check whether the public gateway is reachable without sending a prompt or credential.

Authenticated

GET /v1/models

List HAL virtual model modes and current routing surface.

Authenticated

POST /v1/chat/completions

Use the familiar message format through HAL policy and routing.

Authenticated

GET /v1/peers

Inspect HAL identity and enabled peer nodes without exposing their secrets.

Public smoke

Check reachability without a secret.

curl -fsS https://api.halsupreme.com/health

A healthy response proves the edge answered. It does not prove every provider or model route is ready.

Contribute a peer GPU.

A spare GPU can help HAL only through a bounded, authenticated provider connection. Registry visibility is not routing permission, and a listed node never gains access to the owner’s computer, files, WorkGraph, or credentials.

01Serve one bounded endpointRun an OpenAI-compatible inference server behind HTTPS and authentication. Keep model files, logs, and the node’s secrets on the node.
02Declare the contractProvide model IDs, capabilities, VRAM class, interruption policy, privacy class, rate limit, geographic constraints, and whether the contribution is free, credit-backed, or paid.
03Start disabledThe operator stores the credential in a private environment variable and adds only its variable name to HAL’s peer configuration. No token belongs in a web page, chat message, repository file, or public issue.
04Prove healthHAL checks authenticated model discovery and a tiny, non-sensitive canary. A fresh health and model canary must match the declared capability before admission.
05Admit through ProviderMeshProviderMesh applies cost, privacy, quota, retry, and failure-domain policy. Discovery alone cannot enable a route, and marketplace capacity cannot bypass a zero-dollar budget.
06Run governed jobsHAL sends only an approved workload class, records the selected route and artifact receipt, and disables the peer after stale health, policy mismatch, quota exhaustion, or repeated failure.
What to share privately

Connection metadata, never the credential.

Share the HTTPS base URL, allowed model IDs, capability and privacy classes, expected availability, rate limit, interruption behavior, and a contact for revocation. Send the bearer credential only through the operator’s approved private channel.

Admission state

Disabled until verified.

The normal sequence is candidate → probed → owner reviewed → admitted. A peer may remain useful for discovery without ever becoming routeable.

Workload boundary

No private data by default.

Community and marketplace nodes receive public, synthetic, or separately approved inputs only. Sensitive repository, identity, payment, medical, and unpublished media data stay off the route.

Pick the connection that fits.

Most visitors should use the browser chat. Developers can self-host. Peer access is an intentional operator relationship, not a public-token scavenger hunt.

Talk in the browser

No setup. Choose a mode and start with the composer.

Open chat →

Self-host

Own the gateway, policy, provider configuration, and local credentials.

Open the guide →

Join as a peer

Review the protocol and discuss a bounded, authenticated connection with the community.

Meet the community →