​
Introduction

Malibu turns any Apple Silicon Mac into a remote-addressable, OpenAI-compatible AI inference endpoint.

I'm a provider

Invite-gated pre-beta. Download the app, pick a model, start serving.

I'm a buyer

Get an API key, hit the OpenAI-compatible endpoint.

I'm an agent

Point at get.malibu.tech/skill.md — install, recover, connect.
Malibu is a coordinated inference network built on Apple Silicon — distributed providers, centralized routing and billing. Providers run open-source models on their Macs; buyers hit an OpenAI-compatible API and get responses back with a signed receipt binding (prompt, output, model, provider). Verifiable inference, without a datacenter. What’s live vs planned: Network status.

​
Architecture

Provider Macs connect outbound to wss://coordinator.malibu.tech/ws/provider — no port-forwarding, no static IP, no reverse proxy. Buyers hit https://api.malibu.tech/v1/* and get OpenAI wire-shape responses. Every response also carries an X-Malibu-Receipt header signed with the provider’s Ed25519 receipt key. Public pool snapshot: GET https://api.malibu.tech/v1/stats/overview. Rate card: GET https://api.malibu.tech/v1/rate-card.

​
For providers

Apple Silicon (M1 or newer, macOS 14+) can serve. The Malibu app handles onboarding: invite, model match, GitHub sign-in. New Macs start provisional and may be promoted to pinned. Every billed request produces credits at 90% provider share. USDC-on-Base payout ships at public beta.

​
For buyers

Point any OpenAI SDK client at https://api.malibu.tech/v1 with your Malibu API key. Full OpenAI wire-shape surface: streaming, multi-turn tool calling, JSON-schema structured output, sticky conversations with prefix-cache reuse. Every response carries a signed receipt. List warm models with a Bearer token (GET /v1/models) — the list is small and changes. Console chat at malibu.tech/console includes agent tools, stop, and per-stream decode tok/s. Emitted tool_calls[] are model output. Validate before execution: Buyer-side validation.

​
Trust model

Prompts and responses pass through the gateway to enable routing and billing. Model weights stay on the provider Mac and never leave. Buyer prompts and provider responses are processed as plaintext on provider hardware — providers can technically observe traffic that routes through their machine. This is acceptable for cooperative deployments where buyer and provider have an established trust relationship; it is not a private-inference guarantee. Full trust boundary in Security & trust model.