Classifier Dev

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.

curl https://classifier.dev/spam,not+spam/Win+a+free+iPhone spam

curl "https://classifier.dev/?labels=spam,not+spam&text=Win+a+free+iPhone" # same call, query form spam

Single Cloudflare Worker. No database, no framework, no build step beyond esbuild.

CLI

npm i -g classifier-dev classify bug,feature,praise and lets .github/workflows/publish-cli.yml publish (needs an NPM_TOKEN` repo secret). It talks to the API exactly like curl does.

Layout

src/index.ts routing, validation, tiers, LLM fallback chain, analytics src/query.ts the GET query form, read and written with nuqs; the URL an error suggests src/jev.ts TypeSafe's Jev: packs inputs into requests, reads probabilities src/limiter.ts Durable Object: per-IP rate limiting src/report.ts digest — Analytics Engine SQL -> Resend, flags model fallbacks src/alerts.ts every 15 minutes; emails only when something is wrong src/feedback.ts agent feedback, feedback.now protocol -> email src/privacy.ts keyed pseudonyms: nothing kept points back at a caller src/cost.ts per-request upstream spend, from the providers' own accounting src/docs.ts the site (GET / and GET /benchmark), plain text src/home.ts the same two documents rendered, for browsers only src/ui.ts the shared look: markdown in a terminal cli/ the classify command, published to npm as classifier-dev eval/ benchmarks; read eval/README.md before quoting a number finish-dns.sh one-shot DNS wiring, see below wrangler.example.toml the Worker config, minus the account-specific ids

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:

GET /.well-known/agent-feedback.json what this host accepts GET /api/v1/policy categories, severities, limits POST /api/v1/feedback full structured report POST /api/v1/observations lighter signal POST /api/v1/feedback/{id}/attachments more evidence, later GET /api/v1/receipts/{id} did it land, and was it any good

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.

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:

cp wrangler.example.toml wrangler.toml # once, then fill in your own ids npx wrangler deploy

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:

CLOUDFLAREAPITOKEN dash.cloudflare.com > My Profile > API Tokens > Create Token > "Edit Cloudflare Workers" CLOUDFLAREACCOUNTID the accountid from wrangler.toml STATSKVID the STATS namespace id from wrangler.toml REPORTTO where the daily digest goes

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/001-newsletter-confirmation.sql to the separate newsletter Neon project before deploying. Existing subscribers remain 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: NEWSLETTER_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 database 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), 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/002-newsletter-wants.sql adds the column; 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:

SELECT email, wants FROM subscriber WHERE confirmedat IS NOT NULL AND unsubscribedat IS NULL;

What people asked for first, to order the work by:

SELECT unnest(wants) AS item, count(*) FROM subscriber WHERE unsubscribed_at IS NULL GROUP BY 1 ORDER BY 2 DESC;

The Worker connects as newsletter_writer, which can insert and update confirmation state and the ticks on that one table. Do not read more into that than it deserves: Neon puts every role it creates into neon_superuser, which can read any table whatever the grants say, and neither the project owner nor ALTER ROLE can revoke that membership. So the connection string can in fact read the list. What actually keeps the addresses apart is the separate project and the absent columns, not the grant. To rotate the credential, delete and recreate the role (neonctl roles delete newsletter_writer, then create) and put the new string back with wrangler secret put.

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.

Analytics

Every request writes one Analytics Engine datapoint (tier, label-set fingerprint, country, status, count, latency). No request text is ever stored.

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.

curl -H "authorization: Bearer $REPORTKEY" https://classifier.dev/alerts curl -H "authorization: Bearer $REPORTKEY" "https://classifier.dev/alerts?demo=1&send=1"

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:

curl -H "authorization: Bearer $REPORT_KEY" https://classifier.dev/report

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, labels, caller identifiers or upstream messages are stored.

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:

npx wrangler secret put PRIVACY_SALT # 32 random bytes

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

npm run bench # multi-label, 7 cases: jev vs any OpenRouter model npm run single -- --dataset emotion --backend jev npm run single -- --d

…

Source & license

This open-source MCP server is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: mrmps
• Source: mrmps/classifier-dev
• License: MIT
• Homepage: https://classifier.dev

Install and usage instructions live in the source repository linked above.