The AI Assistant answers questions about the catalog in the context of your account. It knows what every endpoint does, what it costs and what it returns, so it can write a working call instead of a plausible-looking one.
Three things it is good at:
- Generating code. "Validate an email address in Python" comes back as a runnable snippet against the real endpoint, with the parameters it actually takes.
- Finding an endpoint. "I need to verify user signups are legitimate" comes back as a shortlist of endpoints with their credit costs, rather than a name you then have to go look up.
- Debugging. Paste the status and the
errorsentence you got and it will tell you which parameter is wrong, which is usually faster than reading the reference page for a call you thought you had right.
The two modes above the prompt box choose which of the first two you get. Build assumes you know what you want and writes the code. Suggest assumes you know the problem but not the endpoint, and recommends one. Each offers six starter prompts, which are worth a glance the first time to see the shape of question that works well.
The assistant is in alpha. The message and token ceilings below are lower than they will end up being — they exist while the feature is scaled, not as a plan boundary you are meant to buy past.
What comes back
Answers are more than text. The assistant renders:
Endpoint cards, with the logo, description, credit cost and category, plus three buttons — its reference page here in the docs, the same endpoint pre-loaded in the playground, and a mock endpoint created from it. That last one is the fast path from "which endpoint" to "my front end has something to call".
Code blocks with a copy button, in the language you asked for.
Warnings and debug notes, called out rather than buried in a paragraph — usually a gotcha about the endpoint you are about to use.
Anything it recommends can be checked. The endpoint card links to the reference page, and the reference page is generated from the same published definitions the API itself runs on, so a recommendation and the truth cannot drift apart.
Limits
Two separate ceilings, and they are easy to confuse because both end the conversation.
Per conversation: three messages. A conversation is short by design. When you reach the limit the composer says Conversation complete — start a new chat to keep going, and starting a new chat costs nothing. The counter under the box shows where you are.
The practical consequence is worth internalising: ask the whole question in the first message. "Get IP location then fetch local weather, in Node, with error handling" is one message and gets you further than three rounds of narrowing.
Per day: a token allowance, reset at midnight UTC and set by your plan:
| Plan | AI tokens / day |
|---|---|
| Free | 25,000 |
| Starter | 50,000 |
| Pro | 100,000 |
| Mega | 250,000 |
Tokens count the conversation both ways — what you send and what comes back — so a long pasted log costs more of the day's allowance than a short question. The Token usage panel in the sidebar shows what you have spent against the allowance.
When the allowance runs out you get You've reached today's token limit, and the assistant is unavailable until the reset. Nothing else about the account is affected: your keys, your calls and your credits are untouched, because AI tokens and API credits are separate ledgers.
What it does not do
It does not call anything for you. It writes the call; running it is up to you. That separation is deliberate — a suggestion that quietly spent credits would be a surprising bill.
It also has no access to your traffic. It cannot see which endpoints you called or why one failed, so a debugging question needs the actual status and error text pasted in. Where you find those is Analytics, under recent API calls.
Workspaces
The assistant runs against your own account only. Viewing a teammate's workspace through the switcher disables it, with Switch to your own workspace to use the AI Assistant — token allowances belong to an account rather than to a view of one. See teams.
Getting a better answer
The assistant is grounded in the catalog, so questions phrased in terms of the catalog land best.
Name the language and the shape. "In Python, with retries and a timeout" gets you code you can paste; "how do I validate an email" gets you a paragraph.
Describe the problem, not your guess at the solution. In Suggest mode, "I need to stop disposable addresses at signup" gets a better shortlist than "does the email API have a disposable field" — the second forecloses on endpoints that might fit better.
Paste the real error. The error sentence from a failed call names the parameter at fault;
handing that over turns a guessing game into a lookup.
Next
All sources is the catalog it is drawing from.
Playground runs what the assistant writes, and for the same knowledge inside your own editor or agent, the MCP server exposes the catalog as tools.