classifier.dev
Zero-shot text classification. Plain text in, a label and a calibrated
confidence out. No key, no signup. Up to a thousand texts per request.
Single Cloudflare Worker. No database, no framework, no build step beyond esbuild.
CLI
cli/ is a separate npm package (classifier-dev, bin classify): one
dependency-free Node file, tests against a mock API (npm test), semver with
its own CHANGELOG, released with npm run release patch|minor|major which
tags cli-v<version> and lets .github/workflows/publish-cli.yml publish
(needs an NPM_TOKEN repo secret). It talks to the API exactly like curl does.
Layout
The site
curl classifier.dev prints plain text, exactly as it always has. A browser
sends Accept: text/html and gets the same document rendered — headings,
bracketed links, copy buttons — from src/home.ts. Nothing is duplicated: the
page is generated from DOCS and BENCHMARK at request time, so the text
stays canonical and the two cannot drift. ?format=text opts out by hand, and
both responses carry Vary: accept.
Agent feedback
Implements the feedback.now protocol (schema 1.1), so
any agent that speaks it can report a problem without being told how:
Accepted submissions are emailed to REPORT_TO. Reports are kept in KV for 90
days. A repeat of the same domain + surface + category + title is stored and
acknowledged as a duplicate but not emailed again, so one looping agent cannot
empty itself into the inbox; the hourly budget is 100 per IP and the remainder
comes back on every receipt.
Testimonials use signal.category: "testimonial". They require
reporter.agent_type and reporter.agent_description, so praise arrives with
enough context to understand what kind of agent benefited and how it works.
quality_score is a deterministic function of how complete the report is — an
agent can read the rule and write a better one next time. Nothing here calls
the classifier or Analytics Engine: this is where reports arrive saying those
are broken, so it must work when they do not.
Deploy
Merging to main deploys. .github/workflows/deploy.yml typechecks, runs the
Worker and CLI tests, runs wrangler deploy, and then asks the live service for
/v1/health and one classification, so a deploy that uploads a broken Worker
fails in CI rather than in somebody's terminal.
By hand, to try something before it is merged:
wrangler.toml is gitignored and holds the two values that are specific to one
Cloudflare account: account_id, and the STATS KV namespace id that
npx wrangler kv namespace create STATS hands back. The tracked
wrangler.example.toml carries everything else — crons, bindings, migrations —
so the deployment shape is in the repository and only the identifiers are not.
CI has no wrangler.toml, so .github/render-wrangler.mjs writes one from the
example and three repository secrets. That makes the example the deployed shape
rather than a copy of it: change a binding in wrangler.toml alone and CI keeps
deploying the old one.
Repository secrets the deploy needs:
Secrets set with wrangler secret put live on the Worker, not in the script
bundle, so a deploy leaves them alone and CI never needs to know them.
Secrets the Worker reads: TYPESAFE_API_KEY, AI_GATEWAY_API_KEY (Vercel's
AI Gateway, which serves Jev on a free monthly credit; when set it is asked
first and TypeSafe catches what it refuses), OPENROUTER_API_KEY,
CONTEXT_API_KEY (context.dev, the chat's web search and page reads),
RESEND_API_KEY, CF_ANALYTICS_TOKEN, REPORT_KEY, PRIVACY_SALT. Add one
with npx wrangler secret put NAME; none of them are ever read from the
repository. src/index.ts lists the rest in the Env interface.
Secrets are compared with secretEquals (src/secrets.ts), never ===: a
plain comparison returns on the first wrong byte and tells a caller how much
of a guess was right.
The model
Both tiers answer from TypeSafe's Jev, a decision
model rather than a language model: it takes a state and typed questions and
returns a calibrated probability per option, in ~150ms. That shape is why the
API can do three things the LLM version could not.
A thousand inputs per request. State is an array of {id, text} and each
input gets its own question, so the whole batch is one upstream call. The
documented limit is 64k tokens per request; jev.ts packs to a conservative
budget and runs the resulting requests eight at a time. Measured: 400 news
headlines classified in 650ms end to end, and packing 100 items scored the
same as sending them one at a time.
Confidence that means something. On 400 six-way emotion items, answers at
= 0.9 confidence were right 82% of the time and answers below 0.5 were right
29%. The previous model's logprob "confidence" put 87% of news items above 0.9
and was right on 68% of those. So tier: "smart" now means: re-ask the
single-label answers below 0.7 of a fast reasoning model and replace them,
marked escalated: true. Nothing else changes. Which model matters: on
exactly the items Jev is unsure about, deepseek-v4-flash, qwen3.7-flash and
mercury-2.5 were no better than Jev; gemini-3.8-flash took news topics from
87.5% to 90.0% and emotion from 61.8% to 63.7%, so that is the chain. A
frontier model (claude-fable-5.1) gets 72.3% / 90.7% at ~3x the price; the
numbers are on /benchmark if that trade ever looks worth it.
Multi-label in one pass. One yes/no question per label, labels at >= 0.7
returned most-likely-first with the full score map. F1 0.887 on the seven-case
set against 0.799 for the sweep-and-verify LLM cascade it replaced, in 230ms
instead of 1.5s. Re-judging its candidates with the reasoning model made it
worse (and took 23s), so multi-label ignores the tier.
The LLM chains in index.ts remain as the fallback when TypeSafe is
unavailable, limited to twenty inputs because they are one call per input.
The digest reports which model actually answered, with a FALLBACK marker,
because the previous primary was delisted upstream and served its backup for
weeks at F1 0.546 without anything saying so.
Updates list
The form and POST /subscribe send a confirmation email through Resend.
The API returns 202 {"ok":true,"status":"pending_confirmation"}. Nothing is
written to the newsletter database until the emailed token is submitted to
POST /subscribe/confirm as {"token":"..."} (or with the confirmation form).
GET only renders the form, so mail scanners cannot confirm subscriptions.
Tokens are signed with NEWSLETTER_CONFIRMATION_SECRET, expire within 24 hours,
and are never returned from signup. Resend's idempotency key deduplicates repeat
requests for the same inbox within each clock hour. Existing per-IP limits also
apply. Signing-key rotation invalidates outstanding links.
Apply migrations/postgres/ with npm run db:migrate before deploying. The
subscriber table shares the application database. Existing unconfirmed subscribers stay unconfirmed; do not
backfill confirmed_at or send them updates until they confirm. An existing
unsubscribe is never cleared by confirmation or by replaying an old token.
Required Worker secrets: DATABASE_URL, NEWSLETTER_RESEND_API_KEY, and a
random NEWSLETTER_CONFIRMATION_SECRET of at least 32 bytes. NEWSLETTER_FROM
in wrangler.example.toml must use a verified Resend sending domain. REPORT_TO
is the reply address and receives notifications only for newly confirmed rows.
The subscriber table holds email, source, signup/confirmation/unsubscribe dates, which
roadmap items were ticked (wants text[], holding ROADMAP keys; the ticks
ride in the confirmation token and are written on confirmation), the requested
latency when faster inference is selected, and an internal id. It holds no IP,
request id, or classification traffic. Pending
signups are not stored. Tokens and mail-provider error bodies must not be logged.
migrations/postgres/0007_newsletter.sql defines the table; a repeat confirmation
replaces the ticks only when it ticked something, and never clears an
unsubscribe or moves the first confirmation date.
Read only confirmed, active recipients when sending updates:
What people asked for first, to order the work by:
The Worker uses the application DATABASE_URL for newsletter and account data.
Subscriber rows have no account foreign key or API traffic identifier; sharing
storage does not subscribe account holders or change existing consent. The old
newsletter project is retained as a read-only migration archive. See
docs/postgres-setup.md for the verified cutover and recovery procedure.
The copy is one constant — ROADMAP in src/newsletter.ts. The plain text at
curl classifier.dev, the form on the rendered page and index.md all read it,
so a change to the roadmap changes all three or none.
Jev long context
Default Jev and explicit model: "jev" inputs over 32,000 characters use
Chonkie RecursiveChunker with 600-token cl100k_base chunks, parallel Jev
screening for relevant or uncertain evidence (including opposing evidence and
exceptions), then final Jev over whole eligible chunks in source order.
The final evidence budget is 20,000 cl100k_base tokens plus a conservative
provider context check. Eligible chunks can be omitted when that budget fills;
usage.long_context reports selection counts, tokens, calls and timing.
This does not guarantee full-document final reading or universal accuracy.
Use Fast with a workspace key backed by paid balance or an active paid
subscription; anonymous access and signup credit do not qualify. Synchronous limits:
250,000 original context tokens summed across inputs, 20 documents, 32 decisions,
and a 1 MB body. Retail is $0.084/M original cl100k_base context tokens,
counted once across inputs regardless of dimensions and actual inference usage.
Each continuous whitespace or non-whitespace run is limited to 8,192 UTF-16
code units. Longer runs return 400 long_context_input before document
tokenization or chunking, bounding tokenizer work on pathological inputs.
No eligible evidence returns 422 long_context_no_evidence without charge.
Explicit model: "chunklaya" keeps its separate legacy opt-in behavior.
Send one whole document of up to 10 million input tokens (100 MB) to
POST /v1/classify with a funded workspace key. Use the normal JSON input and
labels, or send text/plain with a labels query parameter. Large uploads
return 202 with a status_url; Prefer: respond-async also works for smaller
documents. The server splits, screens, retries and finishes automatically.
The repository CLI handles upload and polling with
node cli/classify.js a,b --document document.txt --json.
The exact whole-document token count sets the reservation and $0.084/M charge.
Source text is stored privately while queued and deleted as it is screened.
Final Jev reads at most 20,000 selected evidence tokens. Completion, failure,
cancellation and 24-hour expiry delete remaining source and evidence; failed
or canceled work is refunded. Results remain available until expiry.
Analytics
Every request writes one Analytics Engine datapoint (tier, label-set fingerprint,
country, status, count, latency). No request text is written to analytics.
Long-context aggregates also go to classifier_long_context_events, with
context/document totals, chunk screening/selection/omission counts, phase input
tokens (including unknown-count flags), answered-call counts and phase timings.
Account events append the corresponding fields; their ordered column contract
is in src/server/analytics/schema.ts.
Documents and original context tokens count once; chunk and phase counters
accumulate across dimension passes. No document text or caller identifiers are
written to the dedicated dataset.
Every request also records what it cost us: OpenRouter returns the charge for
a call when asked, and Jev is billed on the input tokens it reports, at the
rate eval/bench.py prices the benchmarks with. Spend accumulates in a
per-request meter (src/cost.ts) and lands in double3. That column was added
after launch, so it reads 0 for anything older than that deploy.
A cron at 15:00 UTC queries it and emails a digest via Resend.
Alerts
A separate cron runs every fifteen minutes and stays silent unless something
fires. It only watches conditions with an action attached: the Jev key being
refused, Jev not answering (the fallback chain serving quietly, which has
happened), 5xx rates, smart-tier escalations failing (the shape an exhausted
OPENROUTER_API_KEY takes), mean latency, a spend spike against the trailing
day, and traffic stopping outright. 4xx is ignored — that is scanners probing
for /wp-admin, not a fault.
Jev credits. TypeSafe publishes no balance endpoint — its API is
/v1/systemone and /v1/models, nothing else — so there is no number to
watch. Instead the check calls /v1/models with the key every fifteen
minutes and reports back whatever TypeSafe says: a 401, 402 or 403 there means
out of credit, revoked or wrong, and raises a critical alert quoting TypeSafe's
own message rather than guessing which status means what. Because it probes
rather than waiting for traffic, it fires on a quiet host before any caller
meets the fallback chain, and it runs even when Analytics Engine is down.
Each condition emails once when it starts, again every six hours while it
lasts, and once when it clears, with the state in KV under alert:. Thresholds
are the T object at the top of src/alerts.ts.
The first previews without sending or touching state; the second emails a
sample through the real path, to prove delivery works.
Preview the digest any time without sending it:
Jev provider attempts are stored separately in classifier_jev_attempts through
JEV_AE, including recovered failures, retries and gateway cooldown skips.
The report includes status, reason, count and mean latency for each provider;
/alerts shows current incidents even when their notification is suppressed.
The existing 15-minute alert check warns when at least three attempts fail and
failures exceed 5% for either provider. Counts account for Analytics Engine
sampling. No input text, caller identifiers or upstream messages are stored.
Successful simple and multi-label classifier names are retained for 90 days in
a separate aggregate KV record with no caller or source-text fields. Its keyed
fingerprint also appears in per-request analytics, allowing operators to resolve
label names in those pseudonymous records.
AI_GATEWAY_DISABLED = "true" in wrangler.example.toml keeps production on
TypeSafe directly after the gateway repeatedly returned 429 on September 19.
The gateway key is retained. To restore gateway-first routing, verify gateway
capacity, change this variable to "false" in the example and local config,
and deploy. Check provider attempts and the live API tests after re-enabling.
The header is the only way in. A query string lands in access logs, in browser
history and in the Referer header of whatever gets clicked next, so ?key= is
gone. Append ?send=1 to actually email it.
Cloudflare's Analytics Engine SQL is a narrow ClickHouse subset — no uniq(),
no SELECT DISTINCT, and a bare SELECT col ... GROUP BY col is rejected.
Distinct counts therefore use SELECT col, count() ... GROUP BY col and count
the returned rows. Each query is isolated so one failure cannot blank the report.
Privacy
Two columns in that dataset used to be the caller: the IP address, and the
label set, joined and lowercased. Both are keyed hashes now (src/privacy.ts),
so the figures still count distinct callers and distinct classifiers and
nothing can be read back into an address or into somebody's wording. The caller
hash takes the UTC day as well, so it stops being the same value tomorrow —
which is why a unique-caller count over 7d or 30d is really caller-days.
Set the key once, and treat it as a secret like any other:
Rotating it renumbers every fingerprint, so distinct counts double-count across
the rotation. It falls back to ADMIN_SIGNING_KEY, then REPORT_KEY, then a
per-isolate random value, because an unkeyed hash of an IPv4 address or of a
common label set inverts in seconds.
Eval
single.py downloads AG News and dair-ai/emotion test rows on first use and
caches raw results under eval/data/results/ so escalate.py can combine
backends without re-spending. eval/README.md lists the caveats.
Rate limiting
Per IP in a Durable Object, counted in classifications: 3,000/min and
20,000/day on fast, 200/min and 2,000/day on smart.
Two other approaches were tried and rejected:
- Cloudflare's native
ratelimit binding registers fine but never decremented
(70 calls against a limit of 60 all returned success: true).
- KV is edge-cached and eventually consistent, so a counter written this second
is invisible to the next read — every request saw
remaining: 59.
A Durable Object is single-threaded and strongly consistent, which is what a
counter needs. Verified at the original 60/min: 75 requests -> 60 × 200, 15 × 429.
DNS
The domain is registered at Porkbun; the Worker is on Cloudflare. Cloudflare
Workers custom domains require the zone to live in Cloudflare, and neither API
token here has zone.create, so that one step is manual:
- https://dash.cloudflare.com -> Add a domain ->
classifier.dev -> Free plan
./finish-dns.sh — reads the assigned nameservers, points Porkbun at them
via the Porkbun API, and attaches the Worker to the apex and www.
Agent skill
Served from this domain over RFC 8615 well-known discovery, so there is no
repository in the middle:
The index must carry a sha256 of the artifact, and an index that disagrees with
the file makes the skill uninstallable. Rather than commit a digest that a later
edit would silently invalidate, src/skill.ts hashes the bytes it actually
serves, once per isolate. Editing SKILL.md is therefore enough; nothing else
needs updating.
The skill teaches the case the API pitch misses: you are already a model and can
classify anything you can see, so the reason to call out is context, not
capability — filtering forty search results down to six without reading forty.
Two things it documents because testing found them the hard way. Cloudflare
403s Python's stdlib urllib User-Agent before the request reaches the Worker,
so the recipe sets one explicitly. And filters should be told "when in doubt,
keep it": on a ten-snippet research filter that took signal kept from 4/6 to
6/6 with no extra noise, where adding a third "possibly relevant" label did
nothing.
Discovery surfaces
Both are linked from the third paragraph of GET / so an agent reading the
landing page finds them immediately.
Known issue: Cloudflare's managed robots.txt
Adding the zone enabled Cloudflare AI Crawl Control, which prepends a managed
block to /robots.txt disallowing GPTBot, ClaudeBot, CCBot, Google-Extended,
Bytespider, Amazonbot and meta-externalagent. The Worker's own robots.txt is
appended after it and cannot override it.
This blocks training crawlers, not runtime API consumers — any agent can still
call the API. But it does keep the docs out of future model training data, which
works against discovery. The toggles at
dash.cloudflare.com -> classifier.dev -> AI Crawl Control -> Security did not
persist when flipped, so this likely needs a plan-level change or support.
Operator agent access
For operator-owned bulk agent work, set a dedicated AGENT_API_KEY Worker secret
and send it as Authorization: Bearer .... Classification quotas are unmetered,
but inference uses a separate $2/day operator allowance (OPERATOR_DAILY_USD),
shared across IPs, with the same request and concurrency limits as free usage.
It neither consumes nor expands the anonymous allowance and does not replace
ENTERPRISE_API_KEY. The CLI accepts it
through CLASSIFY_API_KEY or CLASSIFIER_API_KEY. It does not authorize private
reports or admin access. Keep it in an ignored secret file; never give it to
public clients. Anonymous quotas continue to apply to unauthenticated traffic.
For a funded workspace key with an arranged bulk quota, set the
API_KEY_RATE_LIMIT_MULTIPLIERS Worker secret to a JSON object mapping the
SHA-256 hex digest of each key to an integer multiplier (1–1,000). Overrides only
raise the admission quota for that key; account counters, balance charging,
model capacity limits, and other keys remain unchanged. Remove the entry to
restore the plan quota.
Finding the classification code
The domain terms are defined in CONTEXT.md.
Both ordinary and dimension classification use the same Jev batching module.
A question group is the smallest part of a request that recovery keeps together:
one input's questions for ordinary classification, one decision for dimensions.
Batch preparation accounts for shared input text once, checks both provider
budgets, and runs before quota charging for dimensions. Execution limits
concurrent requests and splits an oversized batch between question groups.
Change provider budgets and recovery in src/jev.ts; keep dimension meaning in
src/dimensions.ts. Test the posted requests and returned classifications through
jevClassify and the Worker HTTP interface, without depending on packing internals.
Multiple dimensions
POST /v1/classify also accepts items and dimensions:
Each results[i].dimensions[name] carries its own label, confidence, scores,
model and latency. Jev shares state across item–dimension questions, packing
both context limits. Smart escalation is per field and clears the original
scores; unavailable Jev falls back only up to 20 decisions. The API accepts
up to 20 dimensions and 1,000 decisions, with each decision charged to quota.
The classify_dimensions MCP tool uses this same path.
Analytics adds blob9 (single/multi/dimensions), double6 (successful input
items), double7 (dimension count), double8 (uncertain or unscored fields),
and double9 (fields served by fallback). double1 counts successful decisions.
Dimension configurations are keyed fingerprints, never stored verbatim. The
admin panel queries dimension traffic separately, including failures and
caller-days. The existing alert cron flags three or more dimension 5xx in an
hour above 10% of dimension requests, plus any dimension fallback usage.
Run npm run test:e2e with Node 22.18+ against the real production API, or set
CLASSIFIER_BASE_URL to a staging origin. These are named TypeScript tests
using node:test and real HTTP/model calls, with no fetch or inference mocks.
Set CLASSIFIER_API_KEY to an authorized key when the shared IP quota is used
up; it is optional. The suite uses inference and counts toward normal quotas.
The tests check exact decisions and ordering for a 300-decision batch, every
field's allowed labels and probability distribution, real smart escalation,
provider score preservation, aliases, validation errors, MCP, OpenAPI, and
legacy calls.
Responses start as unknown and are validated before becoming typed matrices.
Smart tests require an actual escalation; model drift that makes every fixture
confident fails the test instead of silently skipping the reasoning provider.
npm run typecheck checks the Worker and the TypeScript live tests. Deployment
CI runs the live suite after publishing; npm test stays offline. Fault
injection remains in test/dimensions.test.ts and
test/dimensions-observability.test.ts.
Run the live API tests workflow manually from GitHub Actions (or
gh workflow run e2e.yml) to check production without deploying a Worker.
