Frequently asked questions

A structured file you publish at the root of your site so AI agents can decide whether to recommend your product, ingest your expertise as portable skills, and call your MCP tools.

Same idea as robots.txt (for crawlers) or sitemap.xml (for search engines), shaped for agents. YAML frontmatter plus a markdown body. Open standard, MIT licensed.

Full specification at /docs/spec.

llms.txt is an index of links agents can fetch when researching your site. Useful for pointing agents at documentation they should read. It does not describe your positioning, your audience, your skills, or actions an agent can take.

agents.md describes those things directly, in structured form. The two compose. Most vendors should publish both (the Brief generates both for you).

MCP is the wire protocol an agent uses to call tools and fetch context at runtime. It answers how an agent talks to your services.

agents.md answers a different question: which vendor should the agent recommend in the first place, and whatcan it do on the user’s behalf once it picks one. They compose. An agents.mdfile can declare the vendor’s MCP server endpoint and the tools that will be available when the server goes live.

/agents.md is memorable. Sharable in tweets, in pitch decks, in agent dispatch logs.

/.well-known/agents.mdis the IETF-correct discovery path (RFC 8615). It’s where well-behaved crawlers look first.

Both URLs return byte-identical content. Most setups do this with a single rewrite rule. See the spec for the Next.js example.

Yes. The standard is MIT-licensed and the Zod schema lives at github.com/pretext-md/pretext. Copy the schema file, install zod, validate your file, host at /agents.md. Done.

The hosted Brief is a convenience: it scrapes your homepage, drafts a starting point, and lets you edit. It does not lock you in. You own the file you publish.

Pretext implements the open agents.md standard — we build the hosted Brief, host MCP servers for vendors who want agents to call their tools, and keep the spec and schema open and MIT-licensed.

We think the agent web is the next surface that matters for marketing, and that vendors who show up with structured, honest, action-aware content will get cited and recommended at the expense of vendors who don’t. We’d like to be the company that gives them the tools.

Free forever: the agents.md standard, the spec, the schema, the example files. Use them in any stack, fork them, ignore us.

Free during alpha: the hosted Brief at pretext.md/brief. Generate as many agents.md drafts as you want. Approved alpha users get a per-day rate limit.

Paid (future): hosted Pretext MCP servers. When an agent calls a tool on your behalf (book a demo, fetch a case study, ask a question), Pretext runs the server, the auth, the rate limits. You ship the value, we ship the infrastructure.

The MCP services tab in your Brief output already includes hello@pretext.mdCTAs for build requests. We’re manually quoting these for early customers.

Standards beat features. The convention has to be portable across vendors, across agent platforms, and across years. If we own the standard, we own a bottleneck and the rest of the ecosystem will route around us. So we own none of it.

The spec, the schema, the parser, the conformance tests are all MIT. Our business is the tools on top: the easiest way to generate one, and the only place where the MCP servers are actually hosted for you.

Four artifacts from one URL, in about 15 seconds:

1. agents.md — drafted with your homepage as ground truth, Class V fields marked needs_review so you know what to edit.
2. llms.txt — the companion index of links agents can fetch.
3. MCP suggestions — a list of tools we think agents would want to call on your site, ranked by impact, with a CTA to have Pretext build and host them.
4. Paste payload — a single chunk of text you can drop directly into an AI chat to brief it on your company.

Yes. The Brief draft is just a starting point. Every Class V field arrives marked needs_review precisely because an LLM cannot honestly fill those values from a scrape alone.

Edit until you’d be comfortable with an agent quoting it to a real buyer. Then flip the confidence to vendor_authored and publish.

Today: the Brief fetches your homepage, passes the relevant parts to Anthropic’s Claude with structured-output prompts, and stores the generated draft tied to your account for rate-limiting and history.

We don’t sell your data or share it with third parties beyond what’s required to run the generation (Anthropic for the LLM call, Neon for the database, Resend for the magic link email).

A formal privacy and security page is coming. For now, if you have a specific concern, email hello@pretext.md.

When the Brief hits a CAPTCHA, Cloudflare interstitial, or other bot defense, the output surfaces a blue opportunity panel instead of an error. The panel includes a templated email to hello@pretext.md so we can talk to you directly.

In most cases it’s worth a conversation: agents using browser-style fetches will hit the same wall a buyer’s agent hits. We’ll help you decide whether to allowlist Pretext, allowlist a broader class of agent traffic, or rework the part of your site that’s gated.

A skill is a self-contained instruction kit an agent can load from your agents.md, follow, and apply on behalf of its user. Even offline. Even pasted into a different LLM.

Concretely: a skill has a name (snake_case), a title, a description, a when_to_use hint, the actual instructions (up to 4,000 chars), and optional inputs_needed / outputs. Vendors publish skills to distribute their expertise.

Pretext’s own agents.md ships three example skills: draft_agent_md_for_company, evaluate_agent_readiness, and migrate_from_llms_txt. Read them to see the shape.

Now, for alpha customers. Pretext MCP hosting is live as of Phase 1.5 — every approved alpha org gets a path-based endpoint at https://pretext.md/mcp/<slug> with the ask_vendor tool.

Provisioning is still manual while we onboard the first cohort. Email hello@pretext.md for a slot. Pricing details for the paid tier are on the roadmap — alpha hosting is free.

Pricing isn’t finalized. Early customers will be quoted based on tool complexity, expected request volume, and SLA. Our intent is per-vendor pricing in a familiar SaaS shape (a monthly base + usage), not per-call billing that punishes successful vendors.

The standard, the Brief, and the schema all stay free.

Request access at /signup. We approve accounts manually right now to keep the cohort small enough to actually talk to. You’ll get a magic-link sign-in email once approved.

The bar for approval is low and curiosity-friendly: if you’re a vendor who’d publish your own agents.md, an agent builder consuming agents.md files, or a researcher poking at the standard, you’re in.

For the open standard (schema, spec, examples): file an issue or open a discussion on the public repo.

For everything else (the Brief, account questions, MCP build requests, sales): email hello@pretext.md. Goes to a real human.