HybridAI
Early beta, invite-only API access

One AI endpoint. Private when it matters, fast when it can be.

HybridAI is an OpenAI-compatible gateway. It keeps sensitive prompts on private edge models, sends everything else to the fastest available cloud model, and fails over automatically when a provider is down or rate-limited.

How a request travels

  1. Your app sends a chat requestThe same format as any OpenAI client.
  2. HybridAI checks the promptSensitive-looking prompts take the private edge route.
  3. A model answersIf the first choice fails, the next one takes over automatically.
  4. You get a standard responsePlus headers that show which route and model answered.

What you get

Private by default

Prompts that contain emails, phone-like or card-like numbers, API-key patterns or passwords stay on Cloudflare edge models and are never forwarded to other providers. Detection is pattern-based: a first line of defense, not a guarantee.

Automatic failover

If a provider is rate-limited or retires a model, the next one in the chain answers, and current free models are looked up when the listed ones disappear. Easy prompts use a small, quick model and code gets a bigger one.

See every decision

Each response reports which route and model answered, why, and carries a signed receipt of every provider that saw the request.

Drop-in compatible

Works with OpenAI-style clients. Change the base URL and your key, and you are done.

Real streaming

Words appear as the model writes them, with the same streaming format OpenAI clients already expect.

Saved answers

A repeated or slightly reworded plain question can reuse a saved answer for up to 6 hours, which costs no provider quota.

Live status

See which models are healthy right now, before you build on it.

Clear limits

Keys come in tiers with daily limits that you can read on one page. Nothing is hidden.

More than chat

Voice

Speech to text and text to speech on Cloudflare's own models, with an option to hide private details in the text first.

Hide and restore

Replace emails, phone numbers, cards and keys with placeholders before text goes anywhere, then put the originals back afterwards.

Approvals

Your script asks a person to say yes or no with one tap, and gets a signed answer back.

Webhooks

Your server is told straight away when an approval is decided. Each message is signed so you can check it came from us.

Private memory

Small notes that only your key can read, encrypted so the stored text cannot be read from the database. Searchable by tags.

Dev kit

One-file Python and JavaScript clients for all of this, free to use under the MIT licence.

Control and safety

Agent keys

Give each script or autonomous agent its own key with its own limit. A loop locks only that key, not your main one.

Output guard

Optional: hide or hold back emails, phone numbers, cards and keys that a model writes in its answer.

Bring your own key

Use your own provider keys per request. They are used once and never stored or logged.

Saved endpoint

Save your own provider as the last resort, so you still get an answer when every free provider is busy. Stored encrypted.

Transparency

Which provider answered what share of all requests, as plain counts. No prompts, no people.

Savings meter

An estimate of what your requests would have cost on paid models, on your usage page.

Use it like any OpenAI client

Point your client at the HybridAI address and use the key we send you. The docs have curl, Python and JavaScript examples, streaming, limits and error codes.

const client = new OpenAI({
  baseURL: "https://api.hybridai.cloud/v1",
  apiKey: "YOUR_KEY"
});