InferMux multiplexes the model routers. Point any OpenAI client at one base URL and get the whole catalogue — plus ordered collections that fail over on their own, scoped keys, per-token metering, and a coding agent that runs on our servers and executes on your machine.
curl https://infermux.net/api/v1/chat/completions \
-H "Authorization: Bearer $INFERMUX_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "collection/frontier",
"messages": [{"role": "user", "content": "Ship it."}]
}'Every vendor in the catalogue · refreshed every 3 hours
Everything you would otherwise build twice: fallback, scoping, metering, billing and a catalogue that stays current on its own.
Group models into an ordered collection and call it as collection/your-slug. The order is the routing order: on a failure, a rate limit or an unaffordable model, InferMux falls through to the next member and records every attempt.
Send an image to a mixed collection and the picture goes to the first vision-capable member; its answer is folded back in as a system note and handed to your text model. One request, two specialists.
sk_mr_ keys carry their own model list, collection list, expiry and hard spend cap. Looked up by SHA-256 digest, so auth is O(1) — revoke, expire or cap without touching your account key.
Every upstream attempt writes a row — successes and failures. Charges use the cost the provider actually reported, so your ledger reconciles line by line instead of by estimate.
Subscription allowances count against hour, day and week windows aligned to UTC, not rolling ones — so every reset point is a real instant the API can tell you in advance.
Paste one script tag and your users sign in to mint a scoped, expiring key in their own account. You get BYO-LLM economics without ever holding a password.
Tokens in and out, spend, error rate and every routing attempt — sliced by hour, day, week, month, year and all-time, per key and per model.
Requests carrying tools narrow the candidate list to tool-capable models automatically — 348 of them in the catalogue right now.
Every response carries a router block naming what actually ran: what you asked for, what answered, how many attempts it took and what you were charged. Streaming responses end with the same summary frame.
The request is scoped against your key, priced against the mirrored catalogue and turned into an ordered candidate list.
Candidates are called in order. Failures, rate limits and unaffordable models fall through — each attempt lands in the ledger.
The winning attempt is charged at upstream cost plus markup, or covered by your subscription allowance if the plan includes it.
250 vision models and 314 reasoning models are eligible for staged routing today.
Every screenshot below is a live capture of the running console: the catalogue, collections, the playground, usage, credits and Dave.
$ dave
✓ paired with infermux.net
✓ sandbox: ~/work/api (read/write)
you ▸ find why the webhook retries twice
◦ grep "retry" src/** (local)
◦ read src/webhooks/deliver.ts (local)
◦ edit src/webhooks/deliver.ts (local, approved)
dave ▸ the ack was written after the enqueue —
moved it inside the transaction.
metered as source=dave · covered by planA full agent loop you do not have to host, wired to a client that keeps execution — and your source — on your own machine.
No seat fees, no minimums, no per-model negotiation. You pay what the provider charges plus a flat markup — visible on every model card and every ledger line.
Pay as you go
Prices are stored as the upstream decimal strings, verbatim, and the markup is applied on read — so nothing is rounded away between what the provider charged and what you are billed. Free models stay free: 21 of them consult neither credits nor allowance.
Subscription
Allowance tiers
Token allowances per hour, day and week that cover Dave — and, when the plan allows it, your direct API calls too. Windows reset on UTC boundaries you can predict.
See the tiersCreate an account, mint a scoped key and point your existing OpenAI client athttps://infermux.net/api/v1. That is the whole migration.