Best Tools to Check Whether Coding Agents Can Use Your Docs in 2026

Direct answer

For a free static check of whether coding agents can use your docs, run AFDocs, the open-source audit behind Fern's and Mintlify's scores. To fix what it finds, Mintlify, Fern, ReadMe and GitBook generate llms.txt, Markdown pages and an MCP server. To prove agents can finish a real task from your docs, Lightsage runs 12 coding agents on it.

What are agent-friendly docs?

Agent-friendly docs are documentation a coding agent can find, fetch and read in full, then turn into working code without a person's help. "AI-ready docs" and "docs for agents" name the same goal. Mintlify attributes 78% of October 2026 traffic across the docs it hosts to agents (Mintlify).

The Agent-Friendly Documentation Spec splits the delivery side into 28 checks across 7 categories, as of November 2026. They ask for an index such as llms.txt, Markdown versions of pages, pages small enough for one fetch, stable URLs and no auth gate on public pages. Cloudflare's docs for agents show the result: every page is available as Markdown at /index.md, and each product has its own llms.txt.

Delivery is half of it. The spec's authors say a companion spec on what docs should contain will come later, so a page can pass all 28 checks and still describe a retired endpoint. The other half is agent usability: whether an agent can complete a real task with your product.

Having the files doesn't predict a finished task

Vendors that publish an agent-ready asset score about the same on coding agents' task runs as vendors that don't, in our analysis of Devtool Arena batch 2 (October 2026). Devtool Arena (devtoolarena.com) is Lightsage's public leaderboard of how well 12 coding agents use 150+ developer tools. Each run starts the agent at the vendor's docs URL with nothing else, and scores the task out of 100 (methodology).

We looked at 143 API vendors with 6 or more agent runs, and 139 of them (97%) publish llms.txt.

The gap compares vendors with their category's average, in categories with 4 or more vendors that include both groups. Across vendors with 3 to 7 assets, the average task run stayed between 71.2 and 74.1. The data is observational, groups inside a category are small, and only 4 vendors lack llms.txt, so read it as a pattern.

Mintlify's own test shows what the files do help with. Across 2,400 agent runs on 20 docs sites, agents averaged 2.23 HTTP 404s per task on HTML pages and 0.11 when Markdown pages linked to llms.txt (Mintlify, October 2026). We think both results hold: the index helps an agent find the page faster, and once 97% of vendors have one, the task score depends on what the page tells the agent to do. Mintlify's guide gives the same advice: use readiness scores to find technical issues, then test retrieval with tasks.

How we ranked docs-for-agents tools

These tools do one or more of three things:

1. Generate the files. Build llms.txt, Markdown pages or an MCP server.
2. Check the files. Grade your docs against the Agent-Friendly Documentation Spec or a similar list.
3. Run real agents. Put Claude Code, Codex or other coding agents on your docs and record what they fetched and whether they finished.

We ranked by how much of an agent's view of your docs a tool checks, against what it costs to start in money and setup. A free check that runs on any docs site in a minute ranks above a deeper one that needs a paid plan and tasks you write. A check tied to one platform ranks below one that works anywhere, and tools that build files but check nothing come last. Every vendor fact below comes from that vendor's own pages, read in November 2026.

Site-wide scanners are compared in best agent readiness check tools. For docs, Lighthouse's llms.txt audit flags only a server error and marks a missing file as not applicable. agent-ready.dev's llms.txt validator runs 10 checks against the format.

Docs-for-agents tools at a glance

1. AFDocs: the open-source audit behind the hosted scores

AFDocs is an MIT-licensed CLI and Node.js library that tests a docs site against the Agent-Friendly Documentation Spec. As of November 2026 it runs 28 checks in 7 categories and prints a 0 to 100 scorecard with a fix suggestion per check. You run it with npx afdocs check https://docs.example.com --format scorecard, on Node.js 22 or later.

Its GitHub README says hosted scoring services from Fern, Mintlify, Document360 and Documentation.ai run AFDocs under the hood. It installs as a dev dependency with vitest helpers for CI.

The limitation is that it's static. No agent attempts a task, and the checks cover delivery rather than whether the instructions are right. AFDocs is also 0.x, so check IDs and output formats may change between minor versions.

Price (checked November 2026): free and open source.

Best for: teams on any docs stack, including Docusaurus or a custom site, that want a free spec audit on every build.

2. Mintlify: agent files built in, plus a score and an agent bench

Mintlify is a docs platform that sends every page to AI agents as Markdown, hosts llms.txt, llms-full.txt and a skill.md, and runs an MCP server for each site. Its llms.txt docs say the file stays current with no upkeep.

For checking, mint score grades a public docs site on index links, content negotiation, crawler rules and other readiness items. Mintlify's Agent Score is built on the AFDocs standard. Mintlify also publishes the MIT-licensed Docs URL Discovery Bench, which routes Claude Code and Codex through a proxy while they answer questions about a docs site. It counts 404s, fetches, tokens and time per serving format.

The bench sets the limit. It grades whether an agent found the right page, so working code is outside its scope, and it needs Python 3.9 or later plus Claude Code or the Codex CLI. Agent analytics starts on Pro.

Price (checked November 2026): Starter is free for 5 editors and includes the MCP server. Pro is $450/month billed annually or $540 billed monthly (pricing).

Best for: teams choosing a hosted docs platform that want agent files generated and a way to measure page-finding with real agents.

3. Fern: a free docs plan and a public Agent Score

Fern generates llms.txt at the site root and for every section, product, version and API reference. It prepends an agent directive to every page's Markdown that points agents to .md URLs, llms.txt and the site's MCP server.

Fern's free Agent Score grades any docs URL from 0 to 100. The page describes 22 checks across 7 categories, based on the AFDocs spec. Fern's write-up (April 2026) says failing a foundational check such as discovery or auth access caps the score, and its leaderboard is public.

Agent Score is a static grade, and its 22 checks trail the 28 AFDocs runs as of November 2026. Fern also skips llms-full.txt by design, because its docs say full-site files exceeded most context windows.

Price (checked November 2026): Free is $0 for 10 team members and 1,000 pages, with llms.txt and the MCP server included. Enterprise is custom (pricing).

Best for: API teams that want a free hosted docs plan and a shareable grade to compare with other API docs.

4. Lightsage: real coding agents run your integration task

Lightsage's Agent Usability runs real coding agents on a task you write, such as "verify a webhook signature with our SDK", against your API, CLI or MCP server in sandboxes. Each run returns pass or fail, full traces, errors, and per-agent and per-endpoint scores. The 12 agents are Claude Code, Codex, Cursor, GitHub Copilot, Antigravity, OpenClaw, OpenCode, Hermes, Pi, Kilo Code, Grok and Muse Code.

The Lightsage Docs Agent moves that check into review. It runs your docs evals on each docs pull request, so a change that breaks the integration, such as an SDK page that no longer leads to working code, shows up before you merge.

For the delivery layer, Agent Readiness audits up to 2,500 docs pages on Scale and gives each page a score and a list of issues with suggested fixes. Agent Analytics shows which docs pages coding agents visit.

Agent runs and Agent Readiness need a paid Scale plan, and you write the tasks and pass criteria. Lightsage doesn't host your docs or document generating llms.txt, so your team ships the fixes it suggests.

Price (checked November 2026): Free $0 tracks 10 ChatGPT prompts weekly. Scale starts at $100/month with every coding agent and answer engine in every tier, and Enterprise is custom (pricing). A 7-day trial includes 1,000 credits with no card, and agent runs use usage-based credits.

Best for: devtool teams that need proof that real coding agents can finish their integration from their docs, and want to catch the docs change that breaks it before merge.

5. ReadMe: a grade inside your API hub's settings

ReadMe is an API docs hub that turns on llms.txt, Markdown for Agents and an MCP server from its AI settings. Its AI Discoverability page crawls a sample of your published pages and grades the project against the Agent-Friendly Documentation Spec, on all plans. The same page holds switches for Link headers and an MCP Server Card.

ReadMe also catches agents' wrong guesses. With llms.txt on, a missing .md page returns a Markdown 404 that links to your llms.txt and search URL (LLMs.txt docs).

The limitation is what the grade measures. ReadMe's FAQ says turning on more settings within a group raises that group's score, so the grade mostly shows which switches are off. No agent runs a task.

Price (checked November 2026): Starter is $0 with llms.txt, the MCP server and the AI Dropdown. Pro is $250/month billed annually (pricing).

Best for: API-first teams on ReadMe, or moving to it, that want a built-in grade and agent-friendly 404s without another tool.

6. Context7: see what agents look up in your docs

Context7 indexes library and product docs and serves them to coding agents through its MCP server, CLI and API. As the owner, you can claim your library by hosting a context7.json file with a public key. The admin panel then has a Benchmark tab that evaluates your documentation's quality. Its Metrics tab counts MCP requests from AI coding assistants and the topics developers ask about.

In our Devtool Arena batch 2 analysis, Context7 showed the largest within-category gap of the 7 assets, +4.7 points across 14 categories. We read that as a hint worth testing on your own task.

The limitation is reach. Context7 sees only requests that go through Context7, and agents read its parsed copy of your docs. The pages we read don't document how the Benchmark is scored.

Price (checked November 2026): library owner tools are free. Pro is $10 per seat a month, for private repositories and more API calls (plans).

Best for: library and SDK maintainers who want to see which topics agents look up and steer what Context7 serves them.

7. GitBook: agent and MCP traffic next to human readers

GitBook publishes .md pages, llms.txt, llms-full.txt and an MCP server at /~gitbook/mcp for every published docs site, on every plan including Free (pricing). Its site analytics include an Agents & MCP report on the LLMs, coding agents and crawlers reading your site through Markdown, llms.txt and MCP. The analytics page also shows the search queries agents send over MCP.

A read is a count, though. The pages don't document a task-success signal or a readiness grade. One gotcha: hidden pages stay in llms-full.txt and the MCP server, because hiding a page only removes it from the table of contents.

Price (checked November 2026): Free is $0. Essential is $65 per site a month and Ultimate $249, billed annually, plus $12 per user a month.

Best for: content and DevRel teams that want agent reads and MCP calls counted next to human traffic, on a platform that starts free.

8. Cloudflare Markdown for Agents: Markdown without moving your docs

Markdown for Agents converts your HTML to Markdown at Cloudflare's edge when an agent sends Accept: text/markdown. You turn it on for a zone in AI Crawl Control, or for one docs subdomain or path with a configuration rule.

It suits a home-grown docs site already behind Cloudflare. It converts only HTML, stops at 6 MiB of decompressed HTML and checks nothing, so pair it with AFDocs or a task run.

Price (checked November 2026): included at no cost on Pro, Business and Enterprise plans, and for SSL for SaaS customers.

Best for: teams with docs behind Cloudflare that need Markdown responses without changing their docs stack.

9. llms.txt generators: docusaurus-plugin-llms, Firecrawl's script and agent-ready.dev

docusaurus-plugin-llms (MIT) writes llms.txt and llms-full.txt during a production build, with optional .md copies of each page, and Docusaurus lists it on its community resources page. For any site, Firecrawl's create-llmstxt-py maps your URLs, scrapes Markdown and has GPT-4o-mini write a short title and description per page. agent-ready.dev's generator drafts a file from your sitemap, up to 50 pages, using each page's own title and meta description.

The cost is upkeep. A generated file is a snapshot until you rerun it, and AI-written descriptions need a human read. Firecrawl's script needs Firecrawl and OpenAI API keys and stops at 20 URLs by default.

Price (checked November 2026): the plugin and the script are free and open source.

Best for: teams on Docusaurus or a custom site that need an llms.txt file without changing platforms.

Which should you choose?

Start from the question you need answered.

See also which docs pages coding agents read, MCP server testing tools and how to run agent evals.

How do I write an llms.txt file?

Write a Markdown file at /llms.txt with an H1 naming your product, a one-line blockquote summary, and H2 sections that list links as [name](url): description. The H1 is the only required section in the llms.txt proposal, and an ## Optional section marks links an agent can skip. Point each link at the Markdown version of the page where you have one.

# Example CLI

> Example CLI deploys and manages Example projects from the terminal.

## Docs

- [Install](https://docs.example.com/cli/install.md): Install with Homebrew or npm and log in
- [Command reference](https://docs.example.com/cli/commands.md): Every command, flag and exit code
- [Authentication](https://docs.example.com/cli/auth.md): Use an API key in CI without a browser

## Optional

- [Changelog](https://docs.example.com/cli/changelog.md)

Three habits matter more than the format. The Agent-Friendly Documentation Spec calls an llms.txt under 50,000 characters the single most effective discovery mechanism it observed. Describe what a reader can do on each page, because Fern's write-up notes agents use the description to decide what to fetch. Then test it the way llmstxt.org suggests: ask an agent questions, giving it only your llms.txt as a starting point.

When to choose Lightsage

Choose Lightsage when you need proof that real coding agents can finish your integration from your docs. Most tools on this list generate or check files, and Mintlify's bench measures whether an agent finds the right page. Lightsage runs Claude Code, Codex, Cursor and 9 other coding agents on a real task that starts from your docs. It shows whether each one ends with working code and where it got stuck.

It wins in three situations:

• Your docs pass a spec audit, and agents still fail your integration.
• Your users work in different coding agents, and you need a result for each one.
• Your docs change often, and you want each docs pull request checked before merge.

It's more than you need if all you want is an llms.txt file, which your docs platform or a generator writes for free, or if no coding agent touches your product yet.

Frequently asked questions

What's the best tool to write llms.txt for a dev tools company?

If your docs run on Mintlify, ReadMe, Fern or GitBook, the platform writes llms.txt for you and keeps it current. On Docusaurus, docusaurus-plugin-llms generates it at build time, and for any other site agent-ready.dev's generator or Firecrawl's script drafts one from your pages. Test the result by giving a coding agent only the file and a real question about your product.

How do I audit my docs for AI agents?

Run a spec audit first, such as AFDocs or your docs platform's built-in grade, and fix the failures it lists. Then give at least two coding agents a real task that starts from your docs URL and record where each one stalls. Rerun the same task after every docs change, because a spec grade checks delivery and a content change can break the task without moving the grade.

What's the best tool to make agent-friendly docs for a B2B SaaS company?

A docs platform that generates agent output for you does most of the work: Mintlify, ReadMe, Fern and GitBook all serve llms.txt, Markdown pages and an MCP server. Choose ReadMe for a built-in grade, Fern for a free plan with 1,000 pages, GitBook for agent traffic reports and Mintlify for its benchmark tools. Then check the result with a real task run.

How do I write llms.txt for my CLI?

Use the standard format: an H1 with the CLI's name, a one-line blockquote summary, and H2 sections that link Markdown pages for install, authentication and the command reference. Describe each link by the job it covers, such as authenticating in CI without a browser. Keep the file under 50,000 characters and point to it from the top of every docs page.

Is there software in 2026 that audits docs for agents for my MCP server?

Yes. AFDocs and the hosted grades from Fern, Mintlify and ReadMe audit the docs that explain your MCP server, and ReadMe can also advertise an MCP Server Card. To see whether coding agents can install and use the server from those docs, Lightsage runs Claude Code, Codex and 10 other coding agents against it in sandboxes.