Console · Join as Provider · API Access · Releases

MacProvider

Make any Apple Silicon Mac a remote-addressable MLX inference endpoint. Built on mlx-lm. OpenAI-compatible API — streaming, multi-turn tool calling, JSON-schema structured output, sticky conversations with KV-cache reuse. Every response carries a signed receipt binding (prompt, output, provider, and — with v0.3 receipts — the verified model hash) verifiable with the open-source macprovider-verify CLI. Verifiable inference, without a datacenter.

A lot of the most interesting LLM applications — long-running personal agents, privacy-sensitive tooling, dev workflows that hammer a model thousands of times a day — don't really belong in a cloud datacenter. But the moment you want your Mac's MLX endpoint to be reachable from somewhere that isn't localhost, you fall off a cliff: auth, tunneling, multi-tenant routing, billing, observability, none of it exists out of the box. MacProvider is a thin layer over mlx-lm that fills that gap.

For Providers	For Buyers
Run MLX models locally on any M1+ Mac	OpenAI-compatible /v1/chat/completions endpoint
Outbound WebSocket only — no port-forwarding needed	Route to the full pool or pin to a specific provider
Choose which models to serve from your Mac	Pay only for compute, no subscription required
Manage your provider at portal.streamvc.live	Chat and monitor at console.streamvc.live

Console

console.streamvc.live is the buyer-facing front end. Today it ships two views:

• Browser chat. Send requests to the network directly from the browser, with session history.
• Pool dashboard. Monitor live provider pool status and which models are warm.

portal.streamvc.live is the provider-facing front end. Sign in with your provider_id + provider_token to see your Mac's setup, earnings, identity, and the latest installer release. See SPEC-014 for the surface contract.

Architecture

Provider Mac (MLX)
  └── outbound WS ──▶ MacProvider
                      (pool · routing · billing)
                             │
               ┌─────────────┴─────────────┐
               │                           │
    api.streamvc.live/v1          console.streamvc.live
    (OpenAI-compatible)               (web front end)

New public installs join as provisional providers over outbound WebSocket tunneling and can be promoted to pinned by the operator after observation. No inbound port-forwarding is required on the provider Mac.

For production public onboarding, the coordinator must run with both token gates set explicitly:

auth:
  require_provider_tokens: true
  allow_tokenless_provisional_bootstrap: true

require_provider_tokens keeps normal provider reconnects fail-closed. allow_tokenless_provisional_bootstrap only admits the first tokenless provisional connect far enough to mint and persist its own provider_token; pinned providers and provider IDs whose active token has already been used still reject tokenless reconnects. Invite-only deployments can set the bootstrap flag to false, but then new Macs need an operator-preprovisioned provider_token.

Trust model — what the coordinator sees: 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 prompts and outputs that route 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. A self-hosted coordinator is on the roadmap for buyers who need local-only trust.

For Providers

Run on any Apple Silicon Mac (M1 or newer, macOS 14+):

curl -fsSL https://get.streamvc.live/install.sh | bash

The installer:

• Picks a recommended MLX model by running the SPEC-023 autotune-recommend flow against a signed static feed of demand rank + candidate models. Eligibility is gated by RAM headroom (catalog min_ram_gb vs Mac RAM minus a 4 GB safety margin), bandwidth tier, and a local benchmark probe; scoring ranks eligible rows by demand_weight × completion_rate_per_mtok × sustained_tps using live /v1/rate-card per-token rates (you can override)
• Asks for a stable provider handle used as your pool identity
• Downloads and verifies the latest macprovider-cli release against a signed checksum manifest
• Installs under ~/macprovider and sets up a user-level launchd service
• Runs a local /v1/models check and a coordinator pool visibility check
• Enrolls the provider in autoupdate (SPEC-020): once the coordinator advertises a newer recommended_binary_version, the provider validates the signed release, drains in-flight traffic, and swaps itself in place

Security note: curl | bash gives the downloaded script control of your user account. Inspect first if you prefer:

curl -fsSL https://get.streamvc.live/install.sh -o install.sh
less install.sh
bash install.sh

The binary is checksum-verified against a signed release manifest. macOS quarantine (xattr) is cleared with your approval during install. Developer ID signing and notarization are planned for a future release.

For Buyers

Base URL: https://api.streamvc.live

The API is OpenAI-compatible — swap in your existing client:

from openai import OpenAI

client = OpenAI(
    base_url="https://api.streamvc.live/v1",
    api_key="<your-api-key>",
)

response = client.chat.completions.create(
    model="mlx-community/Llama-3.2-3B-Instruct-4bit",
    messages=[{"role": "user", "content": "Hello"}],
)
print(response.choices[0].message.content)

Get an API key → api.streamvc.live/auth/github/start
API reference → api.streamvc.live/docs#api-reference
Cookbook (tool calling, structured output, sticky conversations, receipt verification) → docs/using-macprovider-with-openai-sdk.md

What the OpenAI-shape surface covers

• Streaming (stream: true) with token-incremental deltas.
• Tool calling (SPEC-018) — multi-turn agent loops for Qwen and Llama 3.3 tool-call templates. First-turn tool_calls[] emission plus second-turn role: "tool" acceptance and assistant-history rendering. Anchor framework: Cline. Caps: 1 MiB per function.arguments, 2 MiB per response, 256 KiB per role: "tool" message.
• Structured output (SPEC-019, LOCKED) — response_format: {"type": "json_schema", "json_schema": {...}} with OpenAI strict-mode subset, and a bundled fix that makes json_object actually enforce top-level JSON.
• Sticky conversations + KV-cache reuse (SPEC-024) — pass a stable conversation_id; the coordinator routes back to the same provider and reports cached_prompt_tokens in the OpenAI usage object. Prefix-cache billing is metered separately.

See docs/using-macprovider-with-openai-sdk.md for worked examples.

Security model: buyer-side validation obligation

For tool-calling responses, emitted tool_calls[] reflect model output, not provider-verified intent; buyer-side agent frameworks MUST validate before execution. MacProvider transports the parsed OpenAI-compatible shape; it does not decide whether a requested tool name or argument payload is safe for your agent policy. Treat emitted tool calls with the same trust posture you would apply to a model running on local hardware: parsed output, not provider-verified intent.

Releases

Latest release and signed binaries: github.com/augustas11/macprovider/releases (badge above auto-updates).

Shipped

• Signed inference receipts (SPEC-015). Every response carries a v0.3 nine-field receipt tuple (model_hash, model_id, output_hash, prompt_hash, provider_pubkey, receipt_version, tokens_out, ttft_ms, unix_ts) signed with the provider's Ed25519 receipt key. Verify with the open-source macprovider-verify CLI; back-compat with legacy v0.1/v0.2 receipts is preserved. Streaming responses now emit a settlement-capable receipt as well.
• Verified model settlement (SPEC-022). Coordinator receipt verdicts drive gateway buyer debit/refund decisions on the money path. Missing / malformed / unverifiable receipts settle as FaultBreakerQualifying with zero provider-positive credits.
• Multi-turn OpenAI tool calling (SPEC-018). First-turn emission plus second-turn role: "tool" acceptance and assistant-history rendering for Qwen and Llama 3.3 templates. Cline drop-in works.
• Structured output (SPEC-019, LOCKED). response_format: json_schema with OpenAI strict-mode subset; post-hoc validate-and-fault.
• Prefix-cache reuse + billing (SPEC-024). Sticky conversation_id routing, cached_prompt_tokens in usage, prefix-cache metering.
• Provider autoupdate (SPEC-020). Coordinator advertises recommended_binary_version; providers auto-drain, verify the signed release, and swap in place.
• Installer-integrated autotune-recommend (SPEC-023 v0.4 LOCKED). Signed static demand-rank + candidate-catalog feeds off Pearl nginx; RAM/benchmark eligibility gates at install time; per-token payout transcript (recommended model or donor-only fallback — no hourly projection or starter tier).
• Provider payout pipeline (SPEC-016). USDC payout design on Base, gated on the operator hot-wallet prerequisites in §9; converged after 20 codex audit rounds.

In flight

• Native Mac app (SPEC-025, "Malibu"). Signed .dmg + menu-bar wrapper around macprovider-cli — non-developer-facing brand and installer path. P0 skeleton shipped.
• Browserless one-click provider onboarding (SPEC-026 draft). Deep-link launch flow so new providers can join without a terminal.
• Self-hosted coordinator. For buyers who need local-only trust; see the trust-model note above.