Docs — AI Function Nodes

Quickstart

Every tool is one POST with a JSON body. There is no session, no handshake and no SDK.

curl -X POST https://YOUR-HOST/api/v1/summarize \
  -H 'content-type: application/json' \
  -d '{"url":"https://example.com","style":"bullets","targetWords":120}'

Tools that produce a file stream the bytes directly:

curl -X POST https://YOUR-HOST/api/v1/markdown-to-pdf \
  -H 'content-type: application/json' \
  -o report.pdf \
  -d '{"markdown":"# Q3 report\n\n- Revenue up 12%","theme":"print"}'

…unless you ask for JSON, which is what you usually want inside an agent:

curl -X POST 'https://YOUR-HOST/api/v1/markdown-to-pdf?output=json' \
  -H 'content-type: application/json' \
  -d '{"markdown":"# Q3 report"}'
# => {"ok":true,"data":{"mime":"application/pdf","filename":"...","bytes":24518,"url":"https://…"}}

Response envelope

JSON responses are always wrapped:

{ "ok": true, "tool": "summarize", "ms": 812, "data": { … } }

Errors use the same shape, and retryable tells an agent whether trying again could help:

{
  "ok": false,
  "tool": "html-to-pdf",
  "error": {
    "code": "quota_exhausted",
    "message": "Browser Rendering is at its free-plan limit…",
    "retryable": true
  }
}

Error codes

Code	HTTP	Meaning	Retryable
invalid_input	400	A parameter failed validation. `details` lists the offending fields.	no
unsupported_media	415	The file type is not supported by the converter.	no
payload_too_large	413	The request body exceeded `MAX_INPUT_BYTES`.	no
rate_limited	429	Per-caller rate limit. Honour the `Retry-After` header.	yes
upstream_blocked	403	The target URL was refused by the SSRF policy (private address, non-http scheme).	no
upstream_failed	502	The target site or an upstream model returned an error.	yes
binding_unavailable	503	A required Cloudflare binding is not configured on this deployment.	no
quota_exhausted	503	The daily free allowance for Browser Rendering or Workers AI is used up.	yes
timeout	504	The page or model took too long.	yes
internal	500	Unexpected server error.	yes

Tool index

Slug (REST)	Name (MCP)	Summary	Cost
html-to-pdf	html_to_pdf	Render raw HTML or a live URL into a print-ready PDF with real Chromium.	browser
screenshot	screenshot	Capture any URL or HTML string as PNG / JPEG / WebP, at any device size.	browser
fetch-page	fetch_page	Retrieve a page as a chosen user-agent — raw HTTP or fully JS-rendered.	free
scrape	scrape	Pull named fields out of a page with CSS selectors — deterministic, no AI tokens.	free
html-to-markdown	html_to_markdown	Strip the chrome, keep the article. Typically 80–90% fewer tokens than raw HTML.	free
document-to-markdown	document_to_markdown	PDF, DOCX, XLSX, CSV, ODT and images in — Markdown out.	AI
markdown-to-html	markdown_to_html	GitHub-flavoured Markdown to a styled, self-contained HTML document.	free
markdown-to-pdf	markdown_to_pdf	One call from model-generated Markdown to a shareable PDF.	browser
summarize	summarize	Compress text or a whole web page into Markdown at a length you choose.	AI
extract-data	extract_data	Turn messy text or a web page into JSON that matches your schema.	AI
translate	translate	Translate text between 100+ languages while keeping Markdown structure intact.	AI
describe-image	describe_image	Read text out of an image, or write alt text and captions for it.	AI

MCP

The endpoint speaks MCP over Streamable HTTP and is completely stateless — POST JSON-RPC, get JSON back. GET /mcp returns 405 by design; there is no server-push channel to open.

{
  "mcpServers": {
    "ai-function-nodes": {
      "type": "http",
      "url": "https://YOUR-HOST/mcp"
    }
  }
}

Supported methods: initialize, tools/list, tools/call, ping, plus empty resources/list and prompts/list so probing clients do not error.

Tool failures come back as isError: true inside a normal result (with a structured error in structuredContent) rather than as JSON-RPC protocol errors, because most clients surface protocol errors as a hard crash instead of something the model can react to.

Function calling

If you are not using MCP, fetch ready-made tool definitions:

# OpenAI / Anthropic style tools[] array
curl https://YOUR-HOST/openai-tools.json

# Full OpenAPI 3.1 document
curl https://YOUR-HOST/openapi.json

# Compact catalog with JSON Schemas and examples
curl https://YOUR-HOST/api/v1/tools

There is also /llms.txt, a plain-text briefing written for a model that has just been pointed at this host.

No login, but shared quota — rate limits and the free tier

Every endpoint here is fully public: no account, no API key, nothing to sign in with. That also means the free-tier allowances below are shared by everyone hitting the deployment, so the only access control is a per-IP rate limit — not identity, just "slow down."

Resource	Free allowance	What we do about it
Browser Rendering	10 min/day, 3 concurrent, 1 new browser / 20s	Warm sessions are reused instead of relaunched; browser tools are limited to a few calls per minute per caller.
Workers AI	10,000 neurons/day	AI tools default to the cheapest capable model; each tool page states its rough cost.
Worker CPU	10 ms / request	HTML parsing is streamed through `HTMLRewriter` rather than buffered into a DOM.
Requests	100,000 / day	Static pages are served by the edge and never invoke the Worker.

When an allowance runs out you get quota_exhausted with a Retry-After header rather than a mystery 500. Check /health for what is currently wired up.

Recipes

Feed a web page to a model without wasting tokens

POST /api/v1/html-to-markdown
{ "url": "https://example.com/long-article", "readability": true, "includeLinks": false }

# data.stats reports the compression, typically 0.10-0.20 of the original HTML

Turn a model's Markdown into a PDF in one call

POST /api/v1/markdown-to-pdf?output=json
{ "markdown": "# Weekly report\n\n…", "theme": "print", "displayHeaderFooter": true,
  "footerTemplate": "<div style='font-size:9px;width:100%;text-align:center'><span class='pageNumber'></span></div>" }

Check whether a site cloaks for search engines

POST /api/v1/fetch-page  { "url": "https://example.com", "userAgent": "chrome-desktop", "include": ["text"] }
POST /api/v1/fetch-page  { "url": "https://example.com", "userAgent": "googlebot",      "include": ["text"] }
# diff the two text bodies

Structured extraction, cheapest path first

# Stable DOM? Deterministic, no neurons:
POST /api/v1/scrape       { "url": "…", "selectors": { "price": ".price", "title": "h1" } }

# Irregular layout? Fall back to the model:
POST /api/v1/extract-data { "url": "…", "fields": ["title", "price:number"] }

Deploy your own

The repository is MIT licensed and deploys to a Cloudflare account in a few minutes — see docs/SETUP-CLOUDFLARE.md. The AI, Browser and rate-limit bindings need no account-specific IDs, so a fresh clone deploys as-is.

Documentation