Jepela accounts are open: a decision API that remembers

Published 2026-10-15 by Jepela

Since 2026-10-15, businesses can sign up for Jepela at app.jepela.com, and Jepela approves each new account.

What Jepela is

Jepela is a hosted decision API. Your code sends a state, such as a support message, and typed questions about it. It gets typed answers back with probabilities, in one call. Nothing is generated: you use Jepela instead of an LLM when your code needs a judgment it can act on, not text it has to parse.

There are three question types, and one request can mix them:

The smallest request is one yes/no question:

POST /v1/systemone
{"state": "Hi, you charged me twice for September. Please send the second payment back.",
 "questions": {"refund": {"type": "noul", "instructions": "Is the customer asking for money back?"}}}

The recorded answer on the examples page gave 0.93 for yes (2026-09-23, jepela-english).

A memory per subject

A subject is whatever you decide about more than once: a customer, a user, a device, a case. Write what you know about it once, as lines. When a request names the subject, Jepela recalls the lines that touch the state, places them before it, and lists the lines it used. With "memory": {"compare": true} the same call also answers without the memory, so you see what it changed.

In the escalation cookbook (2026-09-23), two lines in a priority account's memory moved "escalate to a person?" from 0.20 to 0.82. A control account with no memory stayed at 0.29.

Some things should not be left to a line, so Jepela checks them in code:

• Rules with when compare numbers and dates on the values you send.
• Exclusion rules remove an option before the engine sees it.
• Pinned lines are sent at every decision about the subject.
• forget removes a fact and then checks that recall no longer finds it.
• delete removes a whole subject from memory and disk at once.

What it costs

You pay for the input tokens the engine reads: $0.042 per million. Output is free, because nothing is generated. Writing to memory is free; recalled lines count as input when a decision reads them. The compare, explain and placebo passes are not billed, and neither is a request that fails upstream or in the gateway.

Measured through the API on 2026-09-24 with jepela-english and a 74-character support message:

Credit is prepaid, by card or invoice. Every priced request is one line in your ledger, and GET /v1/usage shows credit, spending and the balance. Jepela is for business customers, and prices exclude VAT.

What it does not do

We publish the engine's limits with the numbers behind them, so you can design around them before you build:

• It does not compare numbers. Limits go in rules with when, checked in code.
• It does not read negation. Naming a forbidden option makes it more likely; use an exclusion rule.
• It reads a window of the state, about 475 tokens on jepela-english (measured 2026-09-22 and 2026-09-23), and says when it cut.
• It reads text only. Describe images in the state.
• English is the most measured language. Other languages are weaker.

Its confidence is the engine's estimate. Whether 0.9 means right nine times in ten is something to measure on your own data, and Jepela's calibration is fitted on your own labels.

How to start

1. Sign up at app.jepela.com. Jepela approves new accounts.
2. Create an API key and add credit by card or invoice.
3. Send a request with curl, the Python SDK or the TypeScript client.
4. Write a few lines with remember, then name the subject and turn on compare.

• Get a key
• Getting started
• Memory
• Pricing
• What the engine gets wrong
• Jepela home