Hayfork · Search as a service
Point Hayfork at a website, upload files, or connect the tools your content lives in. Get a search box that understands questions, a search API for your product, and answers with sources for the AI agents you are building. Nothing to build or tune.
Free plan, no card. Websites and files today.
Confluence, GitHub and Notion next.
Sources
Websites, files, wikis, help desks, code. Hayfork reads each one, keeps the structure that matters, and drops the clutter. Once connected, a wiki page, a web page and a PDF are searched exactly the same way.
Any public site: documentation, help center, blog, changelog. Hayfork reads the content, skips menus and footers, and keeps pages fresh when you re-sync. Sites that publish AI-readable pages index even faster.
PDF, Word, Markdown and text. Drop in a folder of policies, contracts or handbooks. Originals are kept and re-read automatically when we improve how content is understood.
Every connection asks for the least access the source allows, such as one Confluence space or one repository. Credentials are stored encrypted, never shown again, and can be revoked from the console at any time.
Collections
Group sources into collections, and point each search box, product feature or agent at the collection it should cover. The same docs can power the search on your website, the help panel in your app and your support agent, each one scoped to what it needs.
Your workspace is yours alone. Nothing is shared between customers, and every key you hand out does one job. A search key can only search, so the key on your public website can never add, change or delete anything.
Build with it
The same knowledge powers a search box your customers love, answers inside your own product, and the agents your team is building. Add Hayfork once and every agent gets trustworthy answers from exactly the collections you allow.
Understands the question, not just the keywords. Finds the right page whether someone types an exact error code or describes a problem in their own words, and shows where in the document the answer sits.
POST /search { "collection_id": "public-support", "query": "rate limit error 429" }
Ask a question, get back the exact passages that answer it, numbered and linked to their source, sized to fit whatever is asking. Built for in-app help, support tooling and assistants.
POST /retrieve { "collection_id": "support-agent", "query": "how do I rotate an API key", "top_k": 8, "max_tokens": 4000 }
Plug Hayfork into the agent tools your company already uses, so Claude, Cursor and your own agents answer from your knowledge instead of guessing. One MCP server, two tools, your search key. Set it up in a minute.
# any MCP client claude mcp add --transport http hayfork \ https://api.hayfork.dev/mcp \ --header "Authorization: Bearer $KEY"
How it works
Four steps happen behind the scenes. You only do the first one.
Paste a website address, drop in files, or connect a wiki. Re-sync whenever you like. Nothing gets duplicated.
Menus, footers and clutter are removed. Headings and structure are kept, so every answer knows which section of which document it came from.
Content is understood by meaning as well as by exact words, so a precise error code and a vague description both find the right page.
People get clean results with highlights. Agents get the exact passages with sources. Both only ever see the collections they are allowed to.
Use cases
Starting points customers pick most often. Each one is a few sources, a collection, and somewhere to put the search.
Docs, help center, changelog and blog searched together. Customers stop opening five tabs to find one answer, and support stops answering questions the docs already cover.
An Ask AI box on your docs, a support bot that answers from your help center, an assistant inside your product that knows how it works. Each one sees only the collection you give it, and cites what it used.
Help, knowledge base and content search inside your own interface, with results that feel native to your product.
A workspace for each client, their sources connected, a key handed over. One platform behind every assistant you build for them.
Fully hosted
Hayfork runs the whole service for you: reading your sources, keeping them current, and answering every search. There is nothing to install or operate, and your content lives in a workspace that is separate from everyone else's.
Your workspace is separate from everyone else's. Keys are shown once, can expire, and can be revoked instantly.
searchCan only search. Safe on a public website or in an agent's hands.ingestCan add and update content. For your own systems.adminEverything, including creating and revoking keys.Pricing
Start free with your own sources. Upgrade when your collections or your agents need more.
Prices are in US dollars and include tax. A document is a page, file or record, counted once however many collections include it. Re-syncs are free. Searches from people and agents count the same.
Questions
No. Paste the address and Hayfork reads it, keeping the content and dropping menus, footers and widgets. Sites that publish AI-friendly versions of their pages index a little faster, but any site works.
Built-in search covers one property and lives only there. Hayfork indexes many sources into collections you define and serves them anywhere: your website, your product, your support desk, and the agents your team runs.
The exact passages that answer its question, numbered and linked to their source documents, sized to fit the agent's limits. Integrations for popular agent frameworks are available today, and the MCP server connects Claude, Cursor and other agent tools with a search key.
Each connection asks only for what it needs, for example one Confluence space or one repository. Credentials are stored encrypted, never shown again, and can be revoked from the console at any time, taking effect immediately.
Into your own workspace on our servers, separate from every other customer. Basic search never sends it to a third-party AI provider. Premium search, on Team and Business, sends text to Google Cloud to be processed by Google's model, and only once a workspace owner switches it on.
Good enough that we publish the numbers: measured results for both search tiers on public datasets, side by side with keyword search and other search providers, with the method described. In the console you can build your own benchmark from real questions your customers ask and see the score change as you tune it.
Free plan, no card. Start with a website or a folder of files, and have excellent search on it before your coffee is cold.