There's no single winner. The tool that serves AI-agent consumers best depends on which agent-facing surface matters to you: an llms.txt an agent can read before touching your SDK, a generated MCP server it can call directly, or per-operation Markdown docs instead of HTML chrome. Checked directly against each vendor's site as of September 2026: Sourced and Fern both ship real llms.txt files; Stainless does not. Stainless and Speakeasy both generate MCP servers; Fern doesn't appear to. No tool leads on all three at once. The table below is the check, not the verdict — pick rows that match what your agents actually need.
The criteria, and how they were checked
Three observable, agent-relevant facts, each checked directly rather than taken from marketing copy:
- Does
/llms.txtexist and return real content? Checked by requesting the URL directly and reading the response — not the page that talks about agent-readiness. - Is there a documented MCP-server-generation product? Checked against each vendor's own docs pages, not a blog post about MCP in general.
- Does the docs output give agents clean Markdown per operation, or only HTML pages with a chatbot bolted on top?
Comparison table
| Tool | llms.txt (verified Sept 2026) |
MCP server generation | Per-operation Markdown docs | SDK languages | Public self-serve pricing |
|---|---|---|---|---|---|
| Sourced | Yes — generated from the OpenAPI spec, regenerates on every push | Yes — generation plus managed hosting, with multi-host install snippets | Yes — docs preview ships alongside the SDK | TypeScript, Python | Free previews + 2 hosted docs sites + 1 hosted MCP server; paid at publish |
| Stainless | No — stainless.com/llms.txt returns the marketing homepage, not a Markdown file |
Yes — /docs/mcp: a 2-tool "code execution + docs search" architecture, TypeScript only |
Docs Platform is a separate product from SDK generation | ~9, incl. Kotlin, Terraform | Free tier (≤25 endpoints, 5 generators); paid tiers unpriced |
| Speakeasy | Yes — real file at speakeasy.com/llms.txt |
Yes — "Standalone MCP", deploy remote or local | Not confirmed as a distinct per-operation Markdown output | Broad, language-by-language | Free tier is docs-stated (1 SDK, ≤50 methods); pricing page is enterprise-only |
| Fern | Yes — real files at both buildwithfern.com/llms.txt and /learn/llms.txt, with .md-suffix pages per doc |
Not found in Fern's docs as of this check | Yes — .md suffix works on any docs page |
9 production languages | Free forever + Enterprise only, no public middle tier |
| OSS (OpenAPI Generator, openapi-typescript, openapi-python-client) | No — these are codegen tools, not docs platforms | No built-in generator; wire a hand-written MCP server around the generated client with the official MCP SDK | No — you'd host the Markdown yourself | 40+ languages combined across tools | Free, open source |
What each result actually means
Sourced ships all three from one OpenAPI upload — llms.txt, a generated and managed hosted MCP server, and per-operation docs — because they come out of the same generation pipeline rather than three separate products, from one of the last two independent, vendor-neutral hosted SDK-and-docs platforms still standing. The tradeoff is language coverage: TypeScript and Python today, the two ecosystems where teams publish first, against Fern's and Stainless's roughly nine.
Stainless has the most interesting MCP architecture of the four: instead of one tool per operation, it generates two tools — a code-execution tool that runs arbitrary TypeScript against your SDK, and a docs-search tool — on the reasoning that "a small number of tools take up less space in an LLM's context window." That's a real design opinion, not a limitation, but it's TypeScript-only today, and the missing llms.txt means an agent has to already be inside your SDK before that design pays off — there's no equivalent lightweight entry point for one that isn't.
Speakeasy ships a working llms.txt — its content matches the "AI control plane for the AI-native enterprise" positioning documented in the Speakeasy alternatives post — and a standalone MCP product. What isn't publicly priced is what it costs to use past the free tier; see that same post for the enterprise-only pricing page.
Fern is the strongest llms.txt implementation checked here: two working files, .md-suffix Markdown on every docs page, and an explicit "Instructions for AI Agents" block at the top of the file telling agents how to fetch clean text. It's the best machine-readable-docs story of the four. MCP generation, as of this check, isn't part of the product.
The OSS path (OpenAPI Generator, openapi-typescript, openapi-python-client) gives you the widest language coverage and zero vendor lock-in, but none of the agent-facing surfaces — llms.txt, MCP, hosted Markdown docs — come for free. You build and host all three yourself, same as covered in the migration prompt pack for teams comparing a hosted preview against a local generator run.
Which to pick for which scenario
- You want an agent to discover your API before it touches any code. Fern or Sourced — both ship a real
llms.txt, checked directly. - You want an agent to call your API as tools, today, in production. Sourced if you want one managed remote endpoint from the same OpenAPI workflow, Stainless if its TypeScript-only code-execution model fits, or Speakeasy if broader language support matters more than public pricing.
- You want
llms.txt, MCP, and docs from one spec with no separate evaluations. Sourced — see what actually ships in Sourced's/agentsonboarding and the SDK-generation prompt pack for the exact prompts. - You want maximum language coverage and are willing to build the agent layer yourself. OpenAPI Generator or Kiota, then a hand-written MCP server — see what an MCP server actually is before you start.
Honest scope: when this comparison doesn't matter
If your API has no AI-agent consumers today and none on your roadmap, none of this changes your generator choice — pick on language coverage, docs editing, and pricing instead, and the full alternatives guide is the more relevant comparison. And if you're already deep in one of these platforms with a working MCP server or llms.txt you built by hand, a migration purely to consolidate onto one vendor's version is rarely worth the churn — check what's actually missing first.
FAQ
Does Stainless have an llms.txt file?
No, not as of this check in September 2026 — requesting stainless.com/llms.txt returns the marketing homepage (HTML), not a Markdown llms.txt file. Stainless does generate MCP servers, which is a separate capability from llms.txt.
Which of these has the best llms.txt?
Fern's is the most complete checked here: two live files (root and /learn/), explicit agent instructions, and .md-suffix Markdown on every docs page. Sourced's is generated automatically from the OpenAPI spec and regenerates on every push, which trades some hand-tuned polish for staying in sync with the API by default.
Does Fern generate MCP servers?
Not found in Fern's documentation as of this check — its /learn docs index covers SDK generation, interactive docs, and an AI search feature ("Ask Fern"), with no MCP-generation page linked or reachable at the paths checked.
What's the difference between Stainless's MCP approach and a typical one-tool-per-operation server?
Most generated MCP servers, including Sourced's, expose one tool per OpenAPI operation. Stainless instead generates two tools — a code-execution tool that runs TypeScript against your SDK, and a docs-search tool — arguing fewer tools use less of an agent's context window. Both approaches are documented, working designs; which fits depends on how your agent host handles many small tools versus fewer flexible ones.
Can I get llms.txt, MCP, and hosted docs from one free run?
Yes, with Sourced — upload or connect a spec and the preview includes all three, free and unlimited, with up to two hosted noindex docs sites and one hosted MCP server. Create hosted docs from your repo or start free with a direct upload.
Is any of this pricing information final?
Pricing pages change. Every number above is dated to this September 2026 check and cited to the vendor's own docs or pricing page — verify again before budgeting, especially for Stainless and Speakeasy, where paid tiers aren't fully public.
Try it yourself
Run the same OpenAPI spec through Sourced's free preview and compare the generated llms.txt, MCP server, and docs against whatever you're using today — no credit card, and the OpenAPI validator catches spec issues before any of this runs.