The classifier you own.

What you get

How to use it

1. Get an API key

   During the preview, the demo below gives you a key that lasts a day. Account keys arrive with launch.

2. Send labelled examples

   A list of {text, label}, or {text, labels} for multi-label. Add a group (a conversation or document id) so near-copies never land on both sides of the test.

   curlPythonJavaScript

   curl -X POST https://getjex.dev/v1/classifiers \
     -H "Authorization: Bearer $JEX_KEY" -H "Content-Type: application/json" \
     -d '{"name": "tickets", "examples": [
           {"text": "I was charged twice", "label": "billing"},
           {"text": "Export crashes on Safari", "label": "bug"}]}'
   # 202 {"id": "c_…", "status": "queued"}

3. Read the report card

   GET /v1/classifiers/{id} returns the status and, once ready, the report: accuracy and macro-F1 on a held-out test split, per-label precision and recall, the confusion matrix, and the confidence at which it is right 90% of the time.

4. Classify

   curlPythonJavaScript

   curl -X POST https://getjex.dev/v1/classifiers/c_…/classify \
     -H "Authorization: Bearer $JEX_KEY" -H "Content-Type: application/json" \
     -d '{"texts": ["Refund my annual plan"]}'
   # {"results": [{"label": "billing", "confidence": 0.94, "confident": true, "scores": {…}}]}

5. Improve it

   The report lists the test examples it got most wrong. Add examples for the weak labels and train again; each training returns a fresh report so you can compare.

6. Delete it

   DELETE /v1/classifiers/{id} removes the model, its report and any kept examples, and returns a deletion receipt.

Or: classify instantly, with no training examples

Send just your label names (and, better, a one-line description of each). POST /v1/classify answers immediately -- every call, including the very first one on a brand-new label set. The first few calls on a label set you have never sent before come back as mode: "zero_shot_warming" (an instant answer from label-embedding similarity, no training yet); once Jex has generated and fit a small model for that exact label set in the background (seconds, cached from then on), every later call is mode: "instant_head" and more confident. Common label shapes (spam/ham, sentiment, support intents, topics) are pre-warmed, so they never show the cold-start mode at all.

curl -X POST https://getjex.dev/v1/classify \
  -H "Authorization: Bearer $JEX_KEY" -H "Content-Type: application/json" \
  -d '{"labels": {"billing": "a payment, invoice, or charge issue",
                "bug": "something in the product is broken",
                "feature_request": "a request for a new capability"},
       "texts": ["I was charged twice this month for one seat"]}'
# first call on this label set (zero-shot, warming a model in the background):
# {"results": [{"label": "billing", "confidence": 0.4092,
#     "scores": {"billing": 0.4092, "bug": 0.3302, "feature_request": 0.2605}}],
#  "mode": "zero_shot_warming", "ms": 51.0}
# a later call, same label set, once it has warmed:
# {"results": [{"label": "billing", "confidence": 0.804,
#     "scores": {"billing": 0.804, "bug": 0.1375, "feature_request": 0.0585}}],
#  "mode": "instant_head", "ms": 11.8}

import requests
labels = {"billing": "a payment, invoice, or charge issue",
          "bug": "something in the product is broken",
          "feature_request": "a request for a new capability"}
r = requests.post("https://getjex.dev/v1/classify",
    headers={"Authorization": f"Bearer {JEX_KEY}"},
    json={"labels": labels, "texts": ["I was charged twice this month for one seat"]})
j = r.json()
print(j["results"][0]["label"], j["mode"], j["ms"])  # 'mode' is zero_shot_warming at first, instant_head once warm

const labels = { billing: "a payment, invoice, or charge issue",
                 bug: "something in the product is broken",
                 feature_request: "a request for a new capability" }
const res = await fetch("https://getjex.dev/v1/classify", {
  method: "POST",
  headers: { Authorization: `Bearer ${JEX_KEY}`, "Content-Type": "application/json" },
  body: JSON.stringify({ labels, texts: ["I was charged twice this month for one seat"] }),
})
const j = await res.json()
console.log(j.results[0].label, j.mode, j.ms)

2-200 labels, 25+ word-or-two label descriptions recommended for the best cold-start answer. No training examples are billed or stored; generation happens on our side. See the docs for errors, limits and the trained-classifier mode above.

Try it

How Jex compares

Jex vs Jev (by TypeSafe) on 13 public datasets, STORED numbers from a 2026-09-28 run -- no new calls to Jev. Every system saw the same 400 rows from each official test split. Jex's settings were chosen on separate development data before the test was read.

Both instant rows use the same 13 datasets as the table below, measured the same way, 2026-10-01. Second opinion sends only the specific low-confidence texts it checks to an open-model provider (DeepSeek, via our inference gateway) -- off by default, opt-in only. Docs.

Honest picture: Jev leads at no examples or only a handful per label on almost every set, and leads at every size on SST-2 sentiment specifically. Jex catches up around 32-128 examples per label on most sets and leads on most sets with full training data. Instant mode (above) trails this table's "0 examples" column on average (53.8/62.2 vs 58.2) -- it's a different technique (open-weight synthetic generation vs. the original closed-model read) and not directly comparable cell-for-cell; shown honestly as its own line rather than merged into the table.

Latency: Jex p50 12.7 ms server time per classification (measured on the serving CPU, trained/ready-made head, single text); Jev's stored per-item latency on these runs was ~2-230 ms depending on the call shape, not independently re-measured here. Price: Jex $0.005-0.01 per 1,000 texts depending on mode (instant vs. ready-made/custom-trained; second opinion, opt-in, +$0.01/1,000 texts it actually checks); Jev (TypeSafe)'s published list price is $0.042 per million input tokens with output free, which works out to roughly $0.002-0.008 per 1,000 calls depending on text length -- cited from TypeSafe's own pricing, not a new call. Cells show Jex's score over Jev's (accuracy; micro-F1 for GoEmotions). Columns are examples per label; "0" means label names only. Measured 2026-09-28, stored numbers, no new Jev calls. Methodology

Ready-made models

Select by name with {"classifier": "jex/..."} -- no training, no warm-up. Each trained once on licensed public data; numbers measured 2026-10-01.

Full numbers, model cards and licence notes: docs.

Pricing

Privacy

Questions