Web application & API for developers

Omnillink

An address in, ready-to-use data out: an API that writes down its limits and attaches its measurements to every response.

Year
2026
Stack
TypeScript · Node.js · Fastify · Playwright · Redis · Tailwind · Vercel
  • API design
  • Back-end development
  • Application security
  • SDK and developer tools
  • Documentation
  • Presentation site
https://omnillink.kayzen-lyon.com/
Home page of the Omnillink website — Web application & API for developers, built by KAYZEN
routes under /api/v1, described in the OpenAPI 3.1 document
20
read-only tools exposed to AI agents through MCP
8
cookies and audience measurement tools on the site
0
licence of the software, to host yourself
MIT
The starting point

What needed solving.

The background

Omnillink is an API published by KAYZEN from Lyon. You send it the address of a public web page; it returns the page's metadata, a screenshot, a PDF export, custom-extracted fields or its content in Markdown. The software is released under the MIT licence and is self-hosted. The site omnillink.kayzen-lyon.com is its demonstration, with a sandbox and documentation in French and English.

The constraint

Displaying a link preview, archiving a page as a PDF or giving the text of an article to an AI agent calls for the same work every time: reading a page you do not control, sometimes built in JavaScript, sometimes protected, without exposing your own network. The hosted APIs that take care of it charge per call and see the addresses being read. What was needed was a tool you install on your own infrastructure, whose settings and limits are written down, and which cannot be misused to reach an internal network.

The response

One request per need, and an economical reading path: the page is first read by a plain HTTP request, and Chromium is launched only if the result is poor or if the page is an empty application shell. Every address is validated against SSRF attacks before any connection. Every response says what happened: duration, rendering mode, cache status, request identifier and remaining quota. The site follows the same rule: it displays neither volumes nor testimonials, it shows real responses from the API.

Design & art direction

Every design choice, and the reason for it.

  1. The demonstration is on the home page

    The visitor enters an address and launches a real request from the home page, with no key and no sign-up. A developer judges an API by its response: showing it straight away, with the duration, the rendering mode and the cache status, spares them creating an account just to form an opinion.

  2. Each capability is shown with its real output

    The feature cards carry no illustration: each one displays a real response from the API, captured on a test page, and leads to its documentation and to the sandbox already configured. The reader sees the exact format they will receive before writing a line of code.

  3. Evidence rather than figures

    The product has just been released: a volume of requests or a testimonial would have no reality behind it. So the site displays none. Instead, a section lists what anyone can check for themselves: the published security model, the measurements attached to every response, the licence and the limits chosen.

  4. A sandbox that writes the code

    The sandbox gives access to all the parameters and generates the code for the request in eight languages: curl, JavaScript, TypeScript SDK, Python, PHP, Go, Ruby and Rust. The path from trying it out to the first integration comes down to a copy and paste, without reading the whole reference.

  5. The KAYZEN brand identity, light theme and dark theme

    Outfit for headings, Plus Jakarta Sans for text, navy blue and terracotta for the main action: the site takes up the brand identity of kayzen-lyon.com, so that the API is identified as a product of the agency. The light theme, the dark theme or the system theme is chosen in the display options.

Type system

Display

Outfit

Body text

Plus Jakarta Sans

The two families of kayzen-lyon.com, so that the API is recognised as a product of the agency.

Engineering

What's under the bonnet.

  • TypeScript
  • Node.js
  • Fastify
  • Playwright
  • Redis
  • Tailwind
  • Vercel
  1. HTTP first, browser only if necessary

    Launching Chromium for every page costs time and memory. So the page is first read by an HTTP request and analysed; the browser only takes over if the result is poor, if the page is an empty application shell or if the site served a blocking page. The meta.source field indicates the path taken.

  2. SSRF protection on every connection

    An API that visits addresses supplied by the caller can be used to reach an internal network. Private, loopback and cloud metadata addresses are refused, and DNS resolution is pinned to the connection. In the browser, all traffic leaves through an internal proxy that revalidates every connection, redirects included.

  3. No code supplied by the caller is executed

    The API executes no script sent by the client. What precedes a capture is described by parameters and CSS selectors: click, scroll, wait for an element. It is one feature fewer, and that is deliberate: arbitrary code in the browser would contradict the isolation and the SSRF protection.

  4. No cookie or credential passed to the page being read

    The caller can set only three headers: Accept-Language, User-Agent and Referer. Cookie and Authorization are excluded, because responses are cached by address and served to everyone: a page obtained with one caller's session would be served to the next.

  5. Cache, deduplication and a per-host circuit breaker

    Results are cached, in memory or in Redis, and identical requests received at the same time trigger only one processing run. A circuit breaker stops contacting a host after several consecutive network failures. A slow or heavily requested page therefore does not monopolise the available browsers.

  6. An OpenAPI reference derived from the validation code

    The parameters of the OpenAPI 3.1 document are derived from the routes' validation schemas: the published reference describes what the server actually accepts. Every failure carries a stable code, from EINVALIDURL to ETIMEOUT, with the corresponding HTTP status, so that client code can handle it without parsing a message.

  7. An MCP endpoint for AI agents

    Eight read-only tools are exposed on POST /mcp. They go through the same services as the REST routes, and therefore through the same checks. Any result that carries page content begins with a warning: this text comes from the web, it is data to be assessed and not an instruction to follow.

  8. A single function serves the API and the site

    On Vercel, the site is compiled then embedded in the API's bundle: a single function answers everything. The time budget for a request is 55 seconds by default, just under the function's maximum duration (60 seconds), so that the API itself responds with a readable error rather than being cut off by the platform.

Search engine optimisation

How the site makes itself findable.

  1. Sixteen pages, each in French and English

    The sitemap declares 32 addresses: home, sandbox, API reference, ten guides, service status, accessibility and legal notice, in two languages linked by hreflang tags. A developer is looking for a precise answer; so each topic has its own page rather than a section of a single page.

  2. llms.txt files for assistants

    The site publishes /llms.txt, which lists the documentation with a one-line summary per page, and /llms-full.txt, which gives its full text. A coding assistant asked about the API can thus read the up-to-date documentation instead of guessing parameters.

  3. Structured data, and the API kept out of crawling

    The home page declares in JSON-LD the publisher, the API (WebAPI type), the software and its FAQ. The robots.txt file excludes /api/ from crawling: the API's responses are data about other sites, not pages to be indexed under the Omnillink name.

  4. No cookies and no audience measurement

    The site sets no cookies and loads no audience measurement tool; the theme, language and accessibility preferences stay in the browser. So there is no consent banner to get past before reading the documentation.

Accessibility

Usable by everyone.

Since June 2025, the European Accessibility Act has required digital accessibility from a large share of online services. We build it in at the design stage rather than retrofitting it.

  1. An options panel on every page

    A button opens the display settings: text size, enhanced contrast, Atkinson Hyperlegible font, line spacing, underlined links, reduced animations. The choices are saved in the browser and apply to every page, without depending on the system settings.

  2. What has been checked, and what has not

    The Accessibility page lists the checks of 18 November 2026: contrasts calculated for both themes, keyboard navigation in an automated browser, no horizontal scrolling at 375 pixels. It also states what is missing: no test with a screen reader, no WCAG or RGAA audit by a trained person. Saying so avoids suggesting a conformity that has not been established.

Site structure

What was delivered.

  • Home — live demonstration, capabilities, how it works, self-hosting, frequently asked questions
  • Sandbox — all the parameters, code generated in eight languages
  • API reference — routes, parameters, errors, quotas and cache
  • Guides — getting started, self-hosting, SDK, AI agents, recipes, security, limits
  • Service status
  • Accessibility, legal notice and privacy
Method

How the project unfolded.

  1. Core API

    Metadata, screenshot, PDF, extraction and readable content; address validation, cache and deployment as a single function on Vercel.

  2. Hardening

    Internal proxy that validates every browser connection, redirects included; unknown parameters ignored rather than refused.

  3. Developer tools

    TypeScript SDK, command line, MCP server, React component, GitHub action and integration examples.

  4. Site and release

    Sandbox, guides, search, English version, hosted MCP endpoint and changelog; put online in November 2026.

Questions

What we are often asked.

The software is released under the MIT licence: you can use it, modify it and host it yourself, with no cost per request other than your own hosting. The instance that serves the site is a demonstration, with a rate limit.

No. The API responds without authentication. The optional x-api-key header raises the rate limit, and keys are defined by the operator of each instance. By default, the limit is 60 requests per window without a key and 600 with a key.

Public pages over http or https. Private, local or internal addresses are refused by the SSRF protection. Pages built in JavaScript are rendered in Chromium when necessary. If a site responds with a blocking page, the API reports it with an error instead of returning that page as content.

Node.js 22 or later, and Chromium, installed by Playwright. Redis is optional: without it, the cache and the quota counters live in memory, instance by instance. Operation remains the responsibility of whoever hosts it: updates, Chromium, monitoring.

The same high standards, for your website.

A free audit of your current site, a quote within 24 hours and written reasoning — like the case study you have just read. Pay in three interest-free instalments; local funding schemes checked before you sign.

Share