Skip to content

Cloudflare Free Plan

github-actions[bot] edited this page Oct 12, 2026 · 3 revisions

Cloudflare Free plan limits

HelpPuff's default setup runs on Cloudflare's Workers Free plan. These are the limits that matter for it, and what each one means in practice. Checked against Cloudflare's documentation in October 2026. Cloudflare changes them from time to time, so the links go to the current numbers.

Daily limits reset at 00:00 UTC. Workers Paid ($5 a month) lifts almost all of them; the right-hand column says what happens when one runs out.

What one assistant uses

Each assistant (one site) uses 1 Worker, 1 KV namespace, 1 D1 database, 1 Vectorize index, 1 Workflow, 1 Cron Trigger and 1 Durable Object class (live chat's hub, SQLite-backed, the kind the Free plan has; idle, it costs nothing) on your account.

Per account: how many assistants fit

Limit Free For HelpPuff
Cron Triggers 5 per account One per assistant (the daily re-learning). At most 5 assistants on one free account, fewer if your other Workers use crons
D1 databases 10 per account, 5 GB in total One per assistant
Workers 100 per account One per assistant
Vectorize indexes 100 per account One per assistant
Workflows 100 instances running at once, per account Crawls and file uploads queue beyond that; nothing is lost

Per day: how much traffic fits

Limit Free For HelpPuff When it runs out
Workers requests 100,000 a day, per account Every widget request (a chat is a handful), dashboard call and webhook delivery. Static widget files do not count The Worker answers with Cloudflare's error 1027 until midnight UTC
Workers AI 10,000 neurons a day, per account About 30 per answer: roughly 300 answers a day. Crawling costs a few per page, once HelpPuff stops at its own budget first (9,000 by default): visitors get your contact details and a callback form
KV writes 1,000 a day None per message with a database (the default): one per new conversation (the sessions-per-visitor limit), plus settings saves. Without a database, about 4 per message Chats keep working; the sessions-per-visitor limit stops counting until midnight UTC
KV reads 100,000 a day One or two per message (the live config, the conversation's page and form details) Reads fail softly; settings fall back to the deployed ones
D1 rows written 100,000 a day Recording conversations (which the limits and the history read back), leads, learning pages Recording stops for the day; chats continue
D1 rows read 5 million a day Searches, the dashboard Searches fail and the assistant offers a callback
Browser Rendering 10 minutes a day Only pages drawn by JavaScript, while learning the site Those pages are skipped, with the reason shown; the next crawl retries
Durable Objects 100,000 requests and 13,000 GB-s a day [[Live chat Live-Chat]] only: a few requests per live chat (each socket, and one per 20 messages it carries). Hibernating sockets cost no duration. 50 live chats of 40 messages ≈ 250 a day
Vectorize queries 30 million queried dimensions a month About 900 searches a day for a 70-page site Meaning search fails; keyword search still answers

Size

Limit Free For HelpPuff
D1 database size 500 MB per database Conversations, leads and the knowledge base's text. Plenty for a small business: a 70-page site is a few MB, and a conversation a few KB
Vectorize storage 5 million stored dimensions About 4,800 passages at 1024 dimensions (a 70-page site is ~600)
KV storage 1 GB Small: settings, a few counters
D1 Time Travel 7 days How far back helppuff upgrade's database restore point reaches
Uploaded files 10 MB each (HelpPuff's limit) Up to ~300,000 characters of text are learned per file

Per request: why HelpPuff is built the way it is

Limit Free How HelpPuff stays inside it
CPU time 10 ms per request, per Workflow step and per cron run Waiting on AI, the database or the network does not count. Heavy work (crawling, reading files) runs in small Workflow steps
Subrequests 50 to the internet, 1,000 to Cloudflare services, per request The crawl runs in batches of 15 pages per Workflow instance
D1 queries 50 per invocation Writes are batched
Workers AI text generation 300 requests a minute (20 for models that need Workers Paid) Far above a small site's peak
Cron Triggers as often as every minute, in UTC HelpPuff's runs once a day (03:23 UTC)
Workflows 1,024 steps per instance, 1 MiB per step result One step per page; long crawls chain several instances

Signs you have outgrown the Free plan

  • More than a few hundred answers a day (Settings → Knowledge in the dashboard shows today's AI use, and visitors start getting the callback form in the evening).
  • More than 5 assistants on one account.
  • Many pages that need JavaScript rendering.

Workers Paid ($5 a month) removes the daily caps on requests, KV, D1 and Browser Rendering, and bills Workers AI past the free 10,000 neurons at $0.011 per 1,000. Raise backend.budget.dailyNeurons to match what you are happy to spend. See Costs and limits.

Clone this wiki locally