The local-first control plane for coding agents
Every model your agents use. Routed, secured, in view.
One API for your keys and the vendor tools you're signed in to — with fallback chains, per-model cost, and a vault we can't open. Runs on your own machines; no middleman holds your credentials.
One config. Every request in view.
Your API keys and signed-in vendor tools: set models and fallbacks once, saved to your account and followed by every harness and device. Then watch requests, tokens, and cost per model live, all in one dashboard.
One command to install
Install with one command; connect from the dashboard.
Same sk-llm key
Vendor tools you're signed in to join your chains through their official CLIs. No new credentials.
Only on your hardware
The cloud refuses subscription traffic. It runs on machines you own.
OpenLLM usage index
Gateway activity, in the open
What every subscription is worth, which models people actually run, and where the traffic goes — published every period.
Best value
ChatGPT Pro
$42,655 / 30d0%
Claude Max 20x
$15,580 / 30d—
Grok Heavy
$5,778 / 30d—
Busiest providers
Claude
7,499 requests—
ChatGPT
924 requests—
Grok
635 requests—
Know what every model really costs you
OpenLLM privately tracks usage across every key and plan, priced at public list rates, so you can budget, compare models, and right-size what you pay for.
Estimates are informational and may differ from actual pricing. Plan names belong to their vendors; subscriptions stay governed by each vendor's terms, served on your device through official tooling.
An estimate, not a bill
Computed from public list prices. Informational only. It never touches what you owe.
Your account, your data
Built from usage your own accounts report via official tooling. Private to your dashboard, never shared.
Honest both ways
If a plan isn't paying for itself, the card shows that too.
Know what every token costs
Model, tokens, latency, dollars, and which hop actually answered. Metadata only; prompts are never stored.
Every provider, one panel
Paste a key (encrypted before it leaves your browser) or sign in on your daemon. Each card shows what it unlocks and whether it's live.
Fallback chains
Ask for a chain, not a model. Whatever fails, the next hop answers.
Custom endpoints
Any host you can reach, even one on your laptop. Its models route like ours.
Scoped keys
One key per app or device. Its own spend, its own cooldown, revoked alone.
No markup
Your keys, your provider's bill. We charge a flat plan, never a cut.
Built for agents
This page has two readers. You, and your agent.
Everything setup needs — install, provider login, model discovery, the rules an agent must follow — ships as one file it can read: llms.txt. Paste one line into Claude Code, Codex, or Cursor, and the gateway sets itself up.
Read https://www.openllm.sh/llms.txt and set up OpenLLM on this machine.Works in any agent that can read a URL. The guide tells it to ask before anything billable — and never to print your key.
Free to start
50M tokens a month, free
No card, no trial clock. Bring your own provider keys, run 2 devices, and keep every request encrypted end-to-end. Upgrade only when you outgrow it.
Need more? Pro lifts the monthly cap, with a 3-day free trial.
Zero-knowledge, not "trust us"
Keys are AES-256-GCM encrypted in your browser, under a recovery phrase only you hold. We store ciphertext; the decryption key arrives only inside your sk-llm key, lives in memory for the request, then is zeroed.
By design
No master key exists. We couldn't decrypt your vault if subpoenaed.
A breach of our database leaks nothing usable.
Lose the phrase? Reset and re-paste. We can't recover it, by design.
The questions you're actually asking
Can't find what you're looking for? Contact our customer support team