Mintlify Data: Agents Are Reading Your Docs

Someone sent me a link to Mintlify's data page and asked what the product was. There is no product. It is a single public page of charts, and the number on it is more interesting than anything they are selling.

Across every documentation site Mintlify hosts, roughly 65% of requests came from agents rather than humans when I read the page on August 5, 2026. Broken out by industry, crypto and web3 documentation ran 82.5% agents. AI coding tools ran 76.8%. Even ordinary SaaS documentation ran 30%.

If those figures are even roughly right, the audience for your documentation changed while nobody moved the furniture. And the practical consequence is not philosophical. It decides which tools are still viable, because a documentation site that renders in the browser is invisible to a reader that does not run JavaScript.

What the page actually is, and what it is not

Being precise about this matters, because the framing changes how much weight the numbers carry.

mintlify.com/data is a marketing asset. It has no signup, no configuration, no per-customer view, and no product behind it. mintlify.com/docs/data returns a 404. I confirmed the site serves genuine 404s rather than catch-all 200s by requesting a deliberately invented path, so that absence is real rather than a page I failed to find. The page closes with generic "Talk to sales" and "Get started" buttons.

It is also self-measured. Mintlify is reporting statistics about its own fleet, computed by its own method, with no third-party audit and no way for anyone outside the company to reproduce them. That does not make them wrong. It does mean the honest way to cite them is "Mintlify reports," which is how I will use them here.

The method is at least stated plainly on the page:

"Every figure on this page comes from the same request data grouped by human and agent traffic. Human traffic is measured client-side when a browser loads a page. Agent traffic is measured server-side from fetches of machine-readable content and requests from known AI user agents. Figures are rounded to the nearest whole number."

Read that carefully, because the asymmetry is load-bearing. Humans are counted when a browser executes client-side code. Agents are counted server-side. Those are two different measurement systems, and a reader who blocks scripts counts as neither. Any comparison between the two categories inherits that.

One more caution. The counters are live. The absolute request totals moved measurably between two reads a few hours apart on the same day, so I am citing ratios with a date attached and no absolute counts. If you see a specific request total quoted anywhere as a fact, it was already stale when it was written.

The industry split is the part worth keeping

The leaderboard covers what the page describes as "the 150 highest-traffic docs sites, grouped by industry, over the last 30 days." Read on August 5, 2026:

Industry	Agent share	Sites	Requests
Crypto and web3	82.5%	14	23.2M
AI coding tools	76.8%	14	29.9M
Fintech and payments	45.1%	5	1.8M
Dev infrastructure and data	34.1%	39	16.3M
AI models and labs	32.6%	14	56.2M
SaaS and AI apps	30.0%	13	7.3M
Agent platforms and AI tooling	28.1%	24	19.2M

Two things stand out. The spread is enormous, from 82.5% down to 28.1%, so a single fleet-wide average is close to meaningless for planning your own site. And the seven rows sum to 123 sites, not 150, so either 27 sites are unlabelled or the headline percentage and the per-industry rows are computed over different populations. I would not lean hard on any figure that mixes the two.

The counterintuitive row is AI models and labs at 32.6%, well below crypto. The pages most about AI are not the pages most read by AI.

What Mintlify actually sells, and the reporting trap inside it

The product people mean when they say "Mintlify data" is the Analytics feature of the documentation platform. It has six sections: Traffic, Agent views, Assistant, Search, Impressions, and Feedback. Impressions carries an explicit beta warning. Analytics requires a Pro or Enterprise plan.

Here is the part that will quietly mislead you, straight from the docs. Traffic CSV exports carry six columns: humanViews, aiViews, searchIndexViews, trainingViews, otherAiViews, and totalViews. And:

"searchIndexViews, trainingViews, and otherAiViews do not count toward totalViews"

totalViews is only humanViews plus aiViews. So search crawlers, training crawlers and other bots are measured, reported in their own columns, and then excluded from the total. If you glance at totalViews and reason about machine traffic from it, you are understating machine traffic, on the very dashboard you bought to measure machine traffic.

Agent counting is also coarser than the word "visitor" suggests:

"Mintlify identifies agent visitors by IP address and user agent. The agent visitor count approximates distinct sources of AI traffic rather than individual agent sessions or conversations. Multiple requests from the same IP address count as one visitor."

For agents running from a handful of cloud egress addresses, that collapses a great deal of distinct activity into single visitors.

How you actually use it

Setup is genuinely quick if you are comfortable with git. Node 20.17.0 or newer is required.

npm i -g mint          # or: pnpm add -g mint
mint login             # mint signup for a new account
mint new my-docs --name "My Docs" --theme mint
cd my-docs && mint dev --no-open      # local preview on port 3000
mint validate && mint broken-links

Your content is MDX in a git repository, configured by a docs.json file that sets navigation, theme and settings. The site deploys to a .mintlify.site subdomain by default. Connect your own GitHub or GitLab repository from the dashboard's git settings and a GitHub App deploys on push.

Three things that are not obvious from the marketing:

Feedback is off by default and has a hard prerequisite. You enable thumbs ratings, edit suggestions, issue raising and contextual feedback individually on the add-ons page. Edit suggestions and issue raising only work on public repositories. And if you have disabled telemetry in docs.json, you cannot enable feedback features at all.

CSV export arrives by email. You pick one category, and Mintlify emails a download link when the export is ready. It is not a direct download, which matters if you were planning to script around it.

The REST API has seven endpoints and inconsistent pagination. Base URL https://api.mintlify.com, bearer-authenticated with an admin key prefixed mint_, and it requires Pro or Enterprise. The real paths are /v1/analytics/{projectId}/views, /visitors, /searches, /feedback, /feedback/by-page, /assistant, and /assistant/caller-stats. Note that the last three do not match their documentation page slugs, so building URLs from the docs navigation gives you 404s.

curl -H "Authorization: Bearer mint_xxx" \
  "https://api.mintlify.com/v1/analytics/PROJECT_ID/views?dateFrom=2026-07-01&dateTo=2026-08-01&limit=250&offset=0"

Pagination differs per endpoint. views and visitors use offset. searches, feedback and assistant use an opaque cursor. feedback/by-page has neither, only limit plus a hasMore flag. And assistant/caller-stats has no pagination at all. Also, dateTo is an exclusive upper bound: "Results include dates before, but not on, the specified date." An off-by-one day here is easy and silent.

The only documented hourly limit is on key creation, at ten keys per hour per organisation. I found no published per-request rate limit, and I am not going to invent one.

What it costs, and why I am not printing a price

I cannot verify Mintlify's prices, and I am going to say that plainly rather than repeat a number I cannot stand behind.

The pricing page builds every dollar figure out of a rolling digit strip. Each digit position is a stack holding all ten numerals, and a CSS transform slides the right one into view, so if you fetch the page and strip the tags, a price reads $ 0 1 2 3 4 5 6 7 8 9 /mo. There is no number in the markup to read, only the alphabet a number could be drawn from. Pro adds a Monthly and Annual toggle on top of that. Enterprise is the one honest exception and is plain text: "Contact us".

That matters more than it sounds, because it means every Mintlify price circulating on the web was typed by a human squinting at a browser, and none of them can be re-checked by a script. Two figures I was offered while researching this piece, one from a December 2025 blog and one from a mid-2026 tracker, disagreed with each other and with the page's own "Unlimited editor seats" line for Pro, which is the tell for a retired per-seat plan. So I am not repeating any of them. Open the pricing page in a browser and read it with your eyes, and note the date you did it.

What I can verify is the gating, which is often the part that actually decides the question:

Plans are Starter (free), Pro, and Enterprise.
Analytics, feedback and the REST API all require Pro or above.
Enterprise gets "Advanced insights" and analytics streaming to Amazon S3.
Starter has no AI features at all, and you cannot buy credits on it.

The credit economics are documented and worth understanding before you commit. Every Pro plan contributes 10,000 credits per month, overages run at $0.01 per additional credit, and extra credits are sold as recurring monthly add-ons, not one-time top-ups: 15,000 for $145 a month, 40,000 for $370, 90,000 for $800, with 90,000 the largest you can buy without talking to sales. Reading those as one-off purchases understates the annual cost twelvefold, which is an easy mistake to make from the table layout.

There is one contradiction on that page worth knowing about, because it will mislead you in the expensive direction. The pricing FAQ states that "each plan includes a set amount of credits per month." The comparison table on the same page leaves the Starter cell in the Credits row completely empty, while Pro reads "10,000 / month" and Enterprise reads "Custom". Those cannot both be true, and the table is the more specific of the two. I am going with the table, which means you should assume the free tier gets no credit allowance and plan on Pro if you want anything AI shaped. If you are deciding real money on this, that is a question for their sales team rather than either sentence on the page.

The useful thing buried in the same FAQ: you can set a hard spending limit, or switch overages off entirely, from the dashboard. If you turn credits on at all, do that on day one. An uncapped $0.01 per credit against a documentation assistant that agents are hitting is precisely the shape of bill people write blog posts about afterwards.

Two line items on the pricing matrix, "GEO optimizations" and "Agent optimizations", are checked on all three plans and defined nowhere in the documentation. Treat undefined feature names as worth nothing until someone tells you what they do.

The lock-in question, answered specifically

Your prose is MDX in your own git repository. That is the good news and it is real: the words are portable, and no platform can hold them.

What is not portable is everything around them. The docs.json configuration, the MDX component set you used in your pages, the hosting, the API playground behaviour, and your entire analytics history. Leaving means rebuilding the site shell and starting your measurement from zero. The pages themselves come with you.

One detail for anyone with a real API: the interactive playground proxies requests through Mintlify's servers by default (api.playground.proxy defaults to true). You can set it to false, but then requests originate from the reader's browser instead. If your API sits behind an IP allowlist, that default decides whether the playground works at all.

The free GitHub alternatives, verified

I checked every repository below against the GitHub API on August 5, 2026 for licence, archived status and the newest commit on the default branch. None is archived. Thirteen of the fifteen had a commit inside fourteen days; the two exceptions are Nextra at about six weeks and Wiki.js at about three months, and both are called out in place below.

One caution if you run this check yourself, because it is the trap I nearly published. Read the default branch head, not the repository's pushed_at field. pushed_at counts a push to any branch, including pull request branches opened from forks, so a project nobody has actually shipped code to in six weeks still reports a push from yesterday. Nextra is exactly that case: pushed_at said July 31, the default branch said June 23. The licence column is what the repository actually grants, which is not always what people assume.

Static generators. These cost nothing to host, ever, because the output is plain files.

Project	Licence	Stars	The honest catch
Docusaurus	MIT	65.8k	A React app wearing a docs costume. Past config changes you are writing React, and theme "swizzling" is explicitly upgrade-fragile. v3.x is end-of-line, so a v4 migration is already on your horizon. Search means bringing Algolia or a plugin.
Starlight	MIT	9.0k	Still pre-1.0 at 0.41.x, so minor bumps carry breaking changes. Search is built in with no third-party service, which is the real differentiator over Docusaurus.
Material for MkDocs	MIT	27.2k	Read this before choosing it. The maintainers put it into maintenance mode on 2025-11-11 and called 9.7.0 "the final version": critical bug and security fixes only, no new features. The most-recommended OSS docs stack is mid-succession.
Zensical	MIT	5.4k	The Material team's successor project, at version 0.0.53. They publish their own feature-parity gap page against Material. Correct destination eventually, not yet a destination.
VitePress	MIT	18.1k	The release channel is stuck even though the repo is not: stable 1.6.4 has not moved in twelve months while v2 has been in alpha over a year. You sit on a year-old stable or ship an alpha.
Sphinx	BSD-2-Clause	8.0k	reStructuredText by default is a real learning tax. Stock themes look like 2012 unless you add Furo or the PyData theme.
Docsify	MIT	31.4k	Do not use this for public docs in 2026. It renders entirely client-side, so any crawler or agent that does not execute JavaScript sees an empty shell. Given the numbers at the top of this article, that is disqualifying.
Nextra	MIT	13.9k	Static export works but you must opt into it; the default posture is server-rendered. Heaviest toolchain here for the least docs-specific feature set, and no OpenAPI story. Also the quietest project in this table: the newest commit on the default branch is June 23, about six weeks back, and the last release was 4.6.1 in December 2025. Not abandoned, but not moving either.
Fumadocs	MIT	12.8k	Server-first by default, in its own words. The free static path exists but is off by default, which is exactly the trap: the tool that looks cheapest on features quietly wants a server.

For the OpenAPI reference specifically, which is the single thing people genuinely pay Mintlify and ReadMe for: Scalar (MIT, 15.8k stars) renders a spec from one static HTML file you can drop into a page on the site you already own, and Zudoku (MIT, 571 stars) prerenders a full static docs site around a spec. Scalar is a reference renderer rather than a documentation site, so it has no page tree, no prose workflow and no versioning. Zudoku is young enough at 0.x that the config surface still moves and there are few third-party answers when you are stuck late at night.

Wiki-shaped options. These all need a server, a database, and someone to own them. BookStack (MIT, 19.0k) needs PHP and MySQL. Wiki.js (AGPL-3.0, 28.7k) needs Node and Postgres, and note that version 3.0 has never shipped: over a hundred release tags, all of them 2.5.x. Its default branch has also been quiet since May 1, about three months, which is the longest gap of anything I am listing here. Read those two facts together before you put a company wiki on it. Docmost (AGPL-3.0 core, 21.3k) needs Postgres and Redis, and its proprietary tier gates exactly the features a business reaches for first, including multi-factor auth, per-page permissions, API keys, PDF export and audit logs.

Outline (40.0k) deserves a specific warning: it is not open source. It is licensed under the Business Source License 1.1, which permits internal use but forbids operating a competing document service, and does not convert to Apache-2.0 until July 2030. Fine for your own team's docs. A genuine problem if you intended to run it for clients.

Realistic hosting cost for the server-backed group is a 2 vCPU, 4 GB virtual machine, which is $24 a month at DigitalOcean's published price and roughly half that from a European provider. The static group is $0 on Cloudflare Pages, Netlify or GitHub Pages, where GitHub's own documented soft limits are 100 GB of bandwidth a month, a 1 GB site, and ten builds an hour.

Three widely-recommended projects that you should not pick up, because search results still send people to them: VuePress has not been pushed since August 2024, Slate is archived and ships no licence file at all, and Docz is archived. Any tutorial recommending them is old.

When you do not need any of this

For a solo professional with a handful of pages and no API, the cheapest good answer is not a single repository on that list. It is a /docs or /help section on the website you already pay for.

The reasoning is arithmetic, not ideology. A handful of pages is five to twenty pages. You already own a domain, a host and a design. Adding pages costs nothing and introduces no new tooling. Every option above, including the free static ones, adds a repository, a build pipeline, a dependency tree, and a thing that will demand an upgrade in eighteen months. That last item is the actual price, and it never appears in comparison tables. If your site runs on a builder, add pages. If it is already static, add a docs collection and let your docs inherit the design, analytics and search you already have.

Three cases genuinely flip that answer. If you want docs versioned in git and separate from the marketing site, Starlight on Cloudflare Pages is $0 a month and about half a day of work. If you publish an OpenAPI specification, that is the one capability worth paying for or reaching for Scalar. And if you need many editors who will not touch git, a hosted platform is buying you a workflow rather than a website.

Which brings the agent numbers back around. If a growing share of your documentation traffic is machines, the question is no longer which platform has the nicest theme. It is whether your pages are legible without JavaScript, whether machine-readable versions exist, and whether you can see the difference in your own analytics. On that test, a static site on a free host beats a client-rendered one on a paid plan, and Mintlify's own data page is the strongest argument for that position that I have seen a vendor publish.

If you want the wider version of this argument, my book The $20 Dollar Agency makes it at length: most of what small businesses get quoted four figures for is now a config change and an afternoon of attention.

Fact-check notes and sources

All figures and claims verified against official vendor pages and the GitHub API on August 5, 2026. Every claim below survived a second adversarial pass whose only job was to refute the first; several first-pass figures did not survive and were corrected or dropped before publication.

The agent traffic share, the industry leaderboard, and the measurement method: mintlify.com/data. All figures on that page are Mintlify measuring its own hosted fleet, self-reported and not independently auditable, and the counters update live. Cited here as ratios with a retrieval date and attributed to Mintlify rather than stated as established fact.
That the page is not a product: mintlify.com/docs/data returns HTTP 404, and the docs index at llms.txt contains no entry for it. The site was confirmed to serve real 404s rather than catch-all 200s using a deliberately invented path.
Analytics sections, the CSV column definitions, the totalViews exclusion, agent identification by IP and user agent, and the near-real-time delay: Mintlify analytics documentation.
Analytics, feedback and REST API all requiring Pro or Enterprise: same analytics page, plus feedback and API introduction.
Feedback being opt-in, public-repo-only options, and the telemetry prerequisite: Mintlify feedback documentation.
REST endpoint paths, parameters, the exclusive dateTo bound, per-endpoint pagination differences, and the ten-keys-per-hour creation limit: Mintlify analytics OpenAPI specification and API introduction. No per-request rate limit is documented anywhere; do not trust a figure that claims one.
Credit economics, including the recurring monthly add-on tiers and the 90,000 self-serve cap: Mintlify credits documentation.
Plan names and feature gating: Mintlify pricing, read from the page markup on 2026-08-05. Every row of that comparison table tags its cells with an explicit plan name, so the gating above is taken from the markup rather than from reading columns by eye. Worth knowing if you check it yourself: the free tier is the pre-selected column on narrow screens, so a careless read binds the first value it finds to Starter when the value actually belongs to Pro. That is how the credit allowance gets mis-attributed.
The Credits row contradiction: the same pricing page states in its FAQ that "each plan includes a set amount of credits per month," while the Credits row of its own comparison table leaves Starter empty and assigns 10,000 per month to Pro and "Custom" to Enterprise. Both were read from the page markup on 2026-08-05. This article follows the table.
Dollar figures not being machine-readable: each price on that page is a stack of all ten numerals with a CSS transform selecting the visible one, so the markup yields $ 0 1 2 3 4 5 6 7 8 9 /mo rather than a number. That is why no Mintlify price appears in this article. Enterprise reads "Contact us" as plain text.
CLI commands, Node version requirement, MDX and docs.json architecture, and .mintlify.site deployment: Mintlify quickstart.
API playground proxy defaulting to true: Mintlify API playground configuration documentation.
Every repository's licence, archived status, star count and commit recency: queried directly from the GitHub REST API on 2026-08-05. Recency is the committer date of the newest commit on the default branch, from the commits endpoint, not the repository's pushed_at timestamp. The two disagree: pushed_at counts pushes to any branch including fork pull requests, and it reported activity within days for Nextra when the default branch had not moved since June 23. Where GitHub reports a licence as unrecognised, the repository's own LICENSE file was read instead, which is why Outline appears as BSL 1.1 and Sphinx as BSD-2-Clause rather than "Other."
Nextra's last release being 4.6.1 on 2025-12-04: GitHub releases API, confirmed not a prerelease.
Material for MkDocs entering maintenance mode with 9.7.0 as the final feature release: the maintainers' own announcement, dated 2025-11-11.
GitHub Pages soft limits: GitHub's published usage limits documentation.
The $24 monthly figure for a 2 vCPU / 4 GB virtual machine: DigitalOcean's published pricing.

Prices, plan structures and repository health all move. Verify before you commit money, and note that several vendor pricing pages are not readable as text, so a figure you find quoted elsewhere may be old.

Mintlify's Data Page Says Two Thirds of Documentation Traffic Is Now Agents. Here Is What That Costs You.