LLMRegistry connects your app to AI from different companies. One place to send requests, manage costs, and set backup plans.
Think air traffic control. With fewer planes.
And a lot more “explain this like I’m twelve.”
You ask.
Your app sends one request.
We organize.
Models, rules, costs, and backups.
AI answers.
Your app gets the result.
37 small demos. Real concepts. Zero API keys.
Pick a chapter, or let curiosity drive.
Your app sends a request to LLMRegistry. We pass it to the AI model you choose and send the answer back. A model is the AI that does the work. An API is how apps talk to each other.
Press the button. Watch the idea click.
A universal remote for AI. Finally, a remote you cannot lose in the couch.
The gateway uses an OpenAI-style API. Each model still has its own supported features.
These are interactive illustrations with made-up data, timing, and costs. They do not call AI models or change your account.
A handy reference for the “what does that do?” moments.
Open a chapter. Try anything that catches your eye.
No features found. Try “models,” “spend,” or “team.”
Your app sends a request to LLMRegistry. We pass it to the AI model you choose and send the answer back. A model is the AI that does the work. An API is how apps talk to each other.
The gateway uses an OpenAI-style API. Each model still has its own supported features.
Try the demoBrowse AI models by what they can do. Compare their features and listed prices. Pick a model for text, images, speech, or search. Your key and provider connection must allow it.
The live Models page has the current list and prices. These cards are made-up examples.
Try the demoA provider is a company that runs AI models. Connect it with your own provider key, or use shared access when offered. BYOK means “Bring Your Own Key.” Your own key is stored with encryption.
Your provider may bill you directly for BYOK. Shared access and gateway fees follow the current pricing rules.
Try the demoA route is a named list of models and rules. Your app asks for that route. LLMRegistry follows its rules to choose an allowed model. Change the route in one place when your needs change.
Routes use only models you add and that your key can access.
Try the demoTell the route helper what your app needs. It can suggest models, backups, and limits. Review the draft in the route form before you save it. You can also use the form on its own.
The helper needs an admin-selected model and an allowed API key. Helper requests may cost money. A draft is not a saved route.
Try the demoA staged route lets models work in order. One can plan, another can write, and another can check. A repair step can fix a failed check. Your app gets only the final answer.
A route has two to four ordered stages. This works with supported text chat without streaming or tools. Each model call may add cost.
Try the demoWith streaming, an answer arrives in small pieces. Your app can show the first words while the model writes the rest. People have something to read instead of staring at a spinner.
The model must support streaming. Once answer text starts, a backup cannot take over that same answer.
Try the demoAn AI model can ask your app to use a tool, like checking an order. Your app checks the request, runs the tool, and sends back the result. The model can use that fact in its reply.
Your app runs tools and checks access. LLMRegistry carries supported tool requests and results.
Try the demoAsk for a response with set fields, such as item, size, and quantity. JSON is a common format for those fields. Your code can read them without guessing where each piece belongs.
Use a supported model and check the result. A tidy format does not prove the facts are right.
Try the demoAn embedding turns text into a list of numbers that describe its meaning. Your app can compare those lists to find related text. “Comfy shoes” can be close to “soft sneakers.”
LLMRegistry makes the number lists. Your app needs its own database to store and search them.
Try the demoSend a description to a supported image model through an Images route. Your app gets image data or a link. LLMRegistry records the image count and cost.
This demo uses our own drawing. Real results, options, sizes, and prices depend on the image model.
Try the demoSend a supported audio recording through an Audio route. A speech-to-text model turns the spoken words into a written transcript. Your app can show, save, or search that text.
File limits depend on the model. The request record does not save the audio. This demo uses a sample transcript.
Try the demoLaya reads text and makes a small decision. Ask yes or no, pick from a list, or give a score. Use it to sort a support message or decide which team should read it.
The Laya preview has input limits and can make mistakes. Its OpenJev page has recorded examples and a separate live API option.
Try the demoAdd backup models to a route. If the first model hits a problem you selected, LLMRegistry can try the next one. Your app can still get an answer when one model is having a bad day.
Backups work for selected failures before answer text starts. Extra attempts may add cost.
Try the demoFor a time-sensitive text request, ask two models at once. LLMRegistry uses the first valid answer and asks the other model to stop. This can help an answer start sooner.
Both attempts can cost money, even if one is stopped. Racing is an option for supported text routes.
Try the demoWhen a provider keeps failing, a circuit breaker pauses traffic to it. New requests skip that connection. After a short wait, a test checks whether it is ready to work again.
A good test lets traffic return. Another failure keeps the connection paused. Each provider connection has its own health state.
Try the demoPick the lowest model power you will accept. Set the fields you need in an answer. If an answer fails that field check, a stronger model on your list can try again. This works with supported text requests.
Field checks work on supported non-streaming text and skip caching. They check format, not truth. More tries can cost more.
Try the demoThe semantic cache can reuse a saved answer when a new question means the same thing. If the saved answer and rules match, it returns that answer without another model call.
Caching starts off. You choose match and time rules. Only eligible answers in the right project can be reused.
Try the demoA token is a small piece of text. Token Alchemy can remove safe extra spaces in complete JSON and repeated instruction messages. The model gets a shorter copy of the request.
It makes limited changes. Code, tool data, image data, audio, and JSON values stay protected. Savings depend on the request.
Try the demoChoose trusted models at different power levels. Model Matchmaker rates the task and can put a suitable model first. Simple tasks may use a smaller model; harder tasks may need more power.
It picks one allowed model from your route. It does not combine answers or promise that every choice will be cheaper.
Try the demoSet a spending limit for each API key. Before a request starts, LLMRegistry holds an estimate of its cost. If that estimate will not fit, the request stops before a provider does the work.
The hold stops several requests from using the same last dollar. The final charge follows the recorded usage.
Try the demoThe wallet sets money aside before work starts. After the request ends, the ledger records its real cost. Any unused part of the hold returns to the available balance.
Provider costs and LLMRegistry fees depend on your access mode and current pricing. All dollars here are examples.
Try the demoOpen a request record to see its model, provider, time, tokens, and cost. It can show events such as a backup or a cache hit. A request ID helps you find the right record.
The ledger saves request facts, not prompt or answer text. Its timeline may not show every internal step.
Try the demoSee how many requests ran, how many worked, and how long they took. Check each provider’s health over time. Spot a change in the charts. Then open a single request to learn more. In the menu, this is called Observability.
Missing trusted measurements stay blank. Total response time includes more than just time inside the gateway.
Try the demoSee how many requests and tokens your team used and what they cost. Filter by time, model, provider, or project. Export the available CSV file for your own reports.
Use the saved request ledger for exact records. This tour uses sample totals and does not export account data.
Try the demoSee which model got each job and why. The Cosmic Switchboard puts route choices, health, and past traffic on one screen. Open a finished request to see why a model was picked.
It explains recorded choices. Some live stream details are not measured, and prompt text is not included.
Try the demoTest the same prompt on several models or routes. Compare answers, time, tokens, and cost. Add an expected answer or field rules to make the comparison easier.
Real Arena runs use your API key and can cost money. The answers and timing in this demo are fictional.
Try the demoThe public Status page shows known LLMRegistry service problems and updates. You can read it without an account. Check incident notes to see what is affected and what happens next.
This is a fictional incident. A green real status page does not promise every outside model is working.
Try the demoAn API key lets an app call LLMRegistry. Choose which models it can use and how much it can spend. You can add an end date or allowed IP addresses, replace it, or turn it off.
A real key secret is shown once. Store it safely. This demo uses fake passes and never creates a key.
Try the demoKeep your team, keys, routes, and costs in one group, called an organization. Each person gets a role. Viewers can read. Developers manage app tools. Billing handles money. Admins have more control.
Available controls follow each role. Review access when someone joins, leaves, or changes jobs.
Try the demoA passkey lets you sign in with your device or security key. Your device proves it is you with a one-time signed challenge. You can remove old passkeys and end old sign-in sessions.
Your fingerprint, face data, and passkey secret stay on your device. This demo does not start a real sign-in.
Try the demoSet a warning at 75% or 90% of your budget. Add an email address or webhook to receive it. A webhook sends a message to another app. When your saved spend crosses the line, we queue a warning for your team.
Alerts warn you. The hard spending limit blocks new work. Delivery needs a working notification channel.
Try the demoKeep each customer’s work in its own group. Choose their models, budgets, prices, and billing rules. Track what each one used and owes. Make an invoice, or bill, for each customer. These tools need reseller access.
Reseller tools track invoices and payment status. They do not collect payments or send money to sellers.
Try the demoPlatform admins review models before publishing them. They manage allowed platform controls and review an audit log of important changes. The log shows who did what and when.
These controls are for platform admins. Publishing a model still needs working provider, gateway, and user access.
Try the demoSome models have saved scores for how well they work and how much power comes from sources like wind or sun. Cosmic Karma can use those scores to move a good choice up your list. The model must still pass your rules for access, cost, and health.
Energy values are estimates from catalog sources, not live carbon readings. Missing data does not count as clean energy.
Try the demoIf this tool is turned on, platform admins can connect a ChatGPT Codex plan. This opens a separate door for supported coding, chat, and image work. It uses the official Codex tool. Regular members cannot use this door.
This is administrator-only BYOS, or Bring Your Own Subscription. It is separate from ordinary BYOK and shared model routing.
Try the demoAn admin’s Codex plan gets its own session, or saved thread. The next turn can pick up where it left off in the same project folder. Each plan has a separate thread. A temporary turn can skip saved history.
This is for the admin Codex connector. Its runtime keeps thread history. Temporary turns skip the durable session; ordinary API calls do not gain this memory.
Try the demoStart with one model and one request.
Add the extra powers when you need them.