Mivia checks what your AI does, before it does it.
You send text. You get back an answer in a fixed form, a label, or a saved record. When the input is too vague, Mivia says no instead of guessing.
Free 30-day trial, no credit card. The first 1,000 calls are free inside the trial.
Check it
You send an order. Mivia answers in a fixed form, or refuses with the reason. Vague input never becomes a guess.
"Refund my order" → asks: "Which order number?"
Sort it
You show 4 examples per label. Every new text gets your label, with measured accuracy shown before you publish.
"Where is my parcel?" → tracking, 0.91
Save it
Every run is stored as a typed record you can search later. Reads and deletes are never billed.
"Refund #1042, broken on arrival" → record #1042
watch
See it work.
Why this exists
Watch a vague order get stopped.
Three real moments. Every stopped call lands in the record on the right. Pick a scene to step through it.
"Delete the database"
ASK: which one?
001 · vague order · asked back
It says no when it should
Clear runs. Missing asks. Vague stops.
Developers: these are the CALL, CLARIFY, and ABSTAIN verdicts, returned in a fixed form.
Every model call gets checked before it answers. Vague input gets refused, not guessed.
Confident - use it straight away.
The input is complete. The model answers in the exact format you asked for, and your code keeps going.
"Refund order #1042, broken on arrival" → {"intent": "refund"}
Measured, not promised
We test on examples you never showed us.
Every score comes from test examples the model never trained on. Too few examples? We say “not measurable” instead of inventing 99%.
Built-in shapes on the JSON model. Fixed forms, strict output.
The bar a label must clear. Below 0.80 it stays off.
Fixed record types. Search everything as JSON.
Key for REST and MCP. Same shape both ways.
Sort text your way
Text classification API: sort text into your own categories.
Sort text your way. 4 examples per label to start.
Your labels. Your own URL.
Show 4 examples per label. Get your own URL that sorts text. Unpublish or roll back any time.
- See accuracy before you publish
- Below the bar stays unpublished
- Old versions kept
labels ship only above the bar
Every run, kept
Save anything as structured data.
Send raw text, get back a typed record. Search everything later.
Your agent's filing cabinet.
Errors, events, notes, tool calls - stored as typed records across 9 types. Search them any time. Reads and deletes are never billed - only model calls cost.
- Raw text in, typed record out
- Searchable across all 9 types
- Reads and deletes never billed
Pricing, full detail
Three ways to pay, one contract.
Start on a free trial, take the plan after, top up past your 5M.
Try the shapes, the checks, and records before money moves.
- First 1,000 calls free inside trial
- All 12 built-in text types
- Community support
Extra calls, same meter. Credits never expire while your plan runs.
- $10 per 1M past 5M included
- You stay on one plan
- REST and MCP on one key
5M included. Every paid month runs on the plan. Top-ups ride on it.
- 5M calls included monthly
- $10 per 1M overage past 5M
- Priority support
Every endpoint, rate, and limit, written down: read the API docs.
Try the API free for 30 days. No card.
Start free trialAsked often
FAQ: pricing, accuracy, MCP, privacy.
What can call these APIs?
Any HTTP client over REST, and any MCP client: Claude, Cursor, your own agents. Classify answers over REST in v1.
Does it train on my traffic?
No. Requests are processed in memory and discarded. Classifiers train only on examples you paste.
How honest are the classifier scores?
Small sets get re-checked against examples the model never saw. Bigger sets get scored on a separate test set. An empty answer means "not measurable, add examples" - never a made-up 99%.
What happens on bad input?
A typed error, never a fabricated success. Vague calls get refused with the reason attached.
Can we run it in our own cloud?
The model API ships as one CPU-only container with the same methods. Same key format, same behavior.
How much does Mivia cost?
$0 for 30 days, no credit card. Then the $20/mo plan with 5M included, plus top-ups.
What is the classifier ship bar?
A label ships only above 0.80 measured accuracy, shown per label before you publish. Below the bar it stays unpublished.