The Agent-Friendly Web is a design philosophy and ongoing project pursued by Jan Wilmake from 2024 onwards. Its central thesis: the existing internet is built for human eyes — rendered HTML, JavaScript-heavy interfaces, CAPTCHAs, and session cookies — and is deeply hostile to AI agents that consume information as text. Jan's mission is to fix this at the infrastructure level by building URL-accessible, token-efficient mirrors of major information sources.
His own formulation: "I'm On a Mission To Make The Open Internet Agent-Friendly through URLs (Universal Resource Locators)."[1]
Jan built a family of tools — each a mirror of a major platform that returns clean, structured Markdown rather than HTML. These mirrors can be called by AI agents, LLMs, or any HTTP client with no authentication overhead:
| Mirror | Source Platform |
|---|---|
| uithub.com | GitHub repositories |
| googllm.com | Google Search results |
| flaredream.com | Cloudflare Dashboard |
| openapisearch.com | Public APIs (OpenAPI specs) |
| xymake.com | X / Twitter |
| arxivmd.org | ArXiv research papers |
Each mirror follows the same principle: take a URL that would normally return rendered HTML (or require login), and return the same information in a format a language model can consume directly.
The concept is grounded in Jan's broader belief that everything is an information stream and that reducing friction between raw information and AI reasoning is a precondition for building reliable AI agents. A human navigating GitHub can tolerate JavaScript loading states and visual chrome; an LLM calling an API cannot.
Jan connected this to Tim Urban's "Elon's Secret Sauce" framework from Wait But Why (2015): the idea that rigorous first-principles reasoning applied to a continuous information stream produces better outcomes than intuition-based decision-making. UITHub, in this view, is infrastructure for applying that framework at scale.[2]
The deeper commercial insight Jan articulated: rather than building yet another AI application, owning the access layer between the internet and AI systems creates a durable infrastructure position — similar to how search engines created durable positions by indexing the web before anyone else.
Jan identified a recurring difficulty: having built the infrastructure, he struggled to create clear B2B messaging that communicated its value to enterprise buyers. The technology worked; the go-to-market did not.[1]
An alternative framing he explored was a context marketplace — where URL sources would be indexed, curated, and shared across users, with the content provider receiving payment for high-quality indexed content. This remained a design concept as of early 2026.
In October 2025, Jan launched LLMTEXT — an open-source toolkit for the llms.txt standard — under Parallel AI's banner. The toolkit extends the Agent-Friendly Web mission from URL mirrors to documentation standards: rather than mirroring platforms that don't expose clean text, LLMTEXT helps websites publish clean text proactively.
Three tools shipped at launch:
Jan's framing at launch: "AI has overtaken humans as the primary user of the web." The tools respond to data showing an 8% drop in Wikipedia's human traffic as AI summaries absorb queries.[3]
Scalable knowledge work: Jan's 2024 blog post "Scalable Ideas" articulates the same access-layer logic: by automating retrieval of structured information from GitHub (README → landing page pipeline), knowledge work becomes replicable at global scale. The Agent-Friendly Web mirrors are the infrastructure that makes this possible.[4]
The post-smartphone era: Jan's 2024 essay on voice-first AI agents frames open-source agents as the mechanism for eliminating "fences" — proprietary interfaces that prevent machine-to-machine interaction. The Agent-Friendly Web mirrors are, in this reading, fence-removal infrastructure: they convert locked-down HTML surfaces into open text streams.[5]
openapisearch went viral in March 2025 (100k+ pageviews), validating the core premise that developers wanted agent-friendly access to API specifications. uithub had three separate viral events (October 2024, December 2024, March 2025) each exceeding 100k pageviews.[6]
In March 2026, Jan launched Markdown_Browser (markdownbrowser.com) — a browser extension and web tool that shows any URL's markdown representation alongside its llms.txt navigation sidebar. His framing: "The web is bifurcating. For 30 years we had one web — built for human eyes. Now a second web is emerging: markdown, structured data, and llms.txt files that AI agents consume. There's no browser for that second web. Until now."[7]
The Markdown Browser supports pluggable extract engines (Defuddle by default) and pluggable search engines, and was designed to integrate with Parallel AI's API key for rendering HTML-only pages.[8]
Between October 2025 and early 2026, Jan's core thesis that agents would take over as the primary web users received significant external validation:
curl -H 'accept:text/markdown' for the Vercel website[12]Jan noted (October 2025): "Shortly after, @mintlify also adopted it, which probably has a real impact on how this develops as so many startups are using it and now adopt it unknowingly."[9]
Jan has articulated this as a position on open standards: "We need to adopt more open standards, or we lose a shared sense of reality. Markdown is one of them."[13]
In March 2026, Cloudflare launched a /crawl endpoint — one API call to crawl an entire site, returning content in HTML, Markdown, or JSON. Jan's reaction: "Wow". This represented platform-level validation of the Agent-Friendly Web thesis: major infrastructure companies treating markdown output as a first-class concern, not an afterthought.[14]
2025/what-i-want.md ↑2025/the-world-is-information.md ↑../blog/2025/LLMTEXT-for-llmstxt.md ↑../blog/2024/scalable-ideas.md ↑../blog/2024/the-post-smartphone-era.md ↑2025/investors/ryan-hoover.md ↑x-archive/raw/posts/2026-03-02.md ↑x-archive/raw/posts/2026-03-06.md ↑x-archive/raw/posts/2025-10-02.md ↑x-archive/raw/posts/2026-02-12.md ↑x-archive/raw/posts/2026-01-19.md ↑x-archive/raw/posts/2026-01-14.md ↑x-archive/raw/posts/2025-12-08.md ↑x-archive/raw/posts/2026-03-11.md ↑