Local Relay — Serve Models From Your Own Machine

Local relay turns a computer you own into an AnyRouter backend. The anyr CLI opens an outbound connection to AnyRouter, and matching chat-completion requests are forwarded to a local OpenAI-compatible server (Apple fm serve, Ollama, LM Studio, llama.cpp, vLLM). Requests served by your own device are never billed against your credits.

Quickstart

1. Start the relay. The first run signs you in (or reuses your key) and pairs the device.

   anyr relay start

2. Call the model with your normal sk-ar-... key:

   curl https://anyrouter.dev/api/v1/chat/completions \
     -H "Authorization: Bearer $ANYROUTER_API_KEY" \
     -H "Content-Type: application/json" \
     -d '{"model": "apple/foundation-model", "messages": [{"role": "user", "content": "hi"}]}'

The relay starts on-device servers for you: Apple Foundation Models (fm serve on 127.0.0.1:1976) and the PPLX Decider app (127.0.0.1:8080). The status screen marks them (managed); a server you started yourself shows (external) and is left running on quit. Add --no-launch to use only servers already running. Ollama (http://localhost:11434) and LM Studio (http://localhost:1234) are detected, but you start them yourself. Nothing needs to be exposed: the device only dials out, so no port forwarding or public IP.

PPLX Decider on your Mac

perplexity/pplx-decider-flash-v1.2 is Perplexity's on-device decision model. It answers the Decisions API, not chat completions. Only your own Mac serves your own requests: it is never offered to the shared pool, even with --pool.

Install the app from Perplexity's Decider for Mac page into /Applications (AnyRouter hosts no download; the first launch downloads about 5.8 GB). Then run anyr relay start and send a decision request with the model id above. If the app is still downloading, the relay keeps watching and serves it once it answers.

Run anyr relay doctor to check the "Decider app" status:

Donate to the shared pool

anyr relay start --pool

Other users' matching requests can then route to your device when it is idle. Serving earns no credits, but on Go or a higher plan you get free Pool access, as with donating a BYOK key: Pool-served requests cost you $0, even at a $0 balance. This lasts while the device is paired, sharing is on, it was online in the last 7 days, and it offers at least one model the pool can use (Apple's on-device models do not count). Toggle sharing per device, see advertised models and requests served in Dashboard > Devices. Live aggregate availability is on the relay page.

Limits and security

• Apple's on-device model has a hard 4096-token context, shared by instructions, tool schemas and transcript.
• If your device is asleep or offline, AnyRouter detects it within seconds and fails over to your other upstreams for that model. If a pool donor's device serves you instead, normal model pricing applies.
• Pairing mints a one-time-shown token; only its hash is stored. Revoke it in Dashboard > Devices and the device loses access immediately.
• The relay forwards only one fixed chat-completion path. It has no access to your files or shell.
• Only you can route to your device unless you pass --pool; pool routing is anonymous on both sides.