Who reads developer docs now

More and more of the traffic to docs isn't people, and we wanted to know who it actually is. Mintlify hosts docs for thousands of companies, so every time an agent fetches a page, we can see which agent it was. Over the last 90 days we classified hundreds of millions of those reads with agent-traffic, our open-source dataset of agents and crawlers. Here's what we found, and how we measured it so you can check our work.

Claude Code, ChatGPT and Claude do most of the reading

Three agents account for 88% of agent docs reads. Claude Code makes 40% of them. ChatGPT and Codex, which OpenAI sends under the same user agent, make 37%, and Claude, meaning the Claude app and API rather than Claude Code, makes 10%.

If you publish docs, these agents are almost certainly reading yours. Claude, Claude Code and ChatGPT each read about 9 in 10 docs sites on Mintlify in the last 90 days, and all three read 84% of them. Lower down the list, reach and volume start to come apart. OpenCode read 59% of sites but made under 1% of reads, which looks like lots of people using it lightly rather than a few people using it heavily.

New coding agents keep showing up

The list of agents reading docs keeps growing. DeepSeek's harness went from almost nothing in August to 184K docs reads a week, and Kiro grew 13x. OpenCode went from 3K reads a week to 249K, and GitHub Copilot grew 3.9x.

One caveat: part of a jump like DeepSeek's can be an agent starting to identify itself rather than starting to read. More on that in the methodology.

Agents search docs instead of browsing them

Weekly searches through docs MCP servers tripled over the last 90 days, from 262K to 812K, and the number of sites searched each week nearly doubled to about 2,000. The part that surprised us most was how differently agents search. When people search a docs site, the median search is one word, and 88% are one or two words. The median agent search is seven words and reads like a spec, and nearly a quarter run past ten words.

What agents read shows what developers are building

We sorted the 300 docs sites agents read most into eight categories. More than half of agent reads on those sites, 55%, go to docs for AI products, so a lot of what developers build with agents sits on top of other AI. Over the last 90 days, the fastest growth was in creative AI tools (3.4x) and payments (3.3x), as agents wire up more image, video and checkout APIs. Agent frameworks, the biggest category, grew the slowest at 2.0x.

Methodology

Everything here covers August to November 2026, and every number is aggregated across docs sites on Mintlify, so no individual site is named. A few details on how we got there:

• What counts as an agent read. One docs page, as HTML or markdown, fetched by an AI product on behalf of a person. Training crawlers like GPTBot, search indexers like Googlebot and generic HTTP clients don't count, since nobody is waiting on the answer. We also leave out search APIs that fetch pages to build an index, and one site where most of the traffic is an agent reading its own docs.
• How requests are classified. Every request is matched on its user agent with agent-traffic, the same open-source dataset we use internally, so you can check any classification yourself. ChatGPT and Codex send the same ChatGPT-User user agent, so they share a row.
• What we undercount. Agents that fetch through generic tools like curl don't identify themselves, including Codex's shell commands and most of Cursor's traffic, so their real totals are higher than shown. An agent's first appearance is when it started identifying itself, which can be later than when it started reading.
• Which sites count. Shares cover every agent read. Reach only counts the 4,935 docs sites with at least 2,000 browser page views, which leaves out empty, test and spam sites.
• How growth is measured. We compare the first and last full week of the window. Category growth only counts sites with coding-agent traffic in both weeks, so sites that joined or left Mintlify partway through don't skew it.
• Categories. We hand-sorted the 300 docs sites agents read most into eight categories, based on what each product does.
• Searches. An agent search is one call to the search tool on a docs site's MCP server. For people, we use the query they clicked a result from in the docs search bar, so half-typed queries don't count.