Swaru

Help

Your knowledge base

One set of articles for your team, your chat and any agent you connect.

Articles live on the Articles page. Published articles are what your chat answers from, what your public help pages show, and what agents read. Agents can suggest changes, but only someone on your team can publish one.

Bring your articles over

An admin does this on Articles, under Import. Nothing is sent back to the old tool, and nothing keeps the two in step afterwards. Every article lands as a draft, so nothing changes for your visitors until you publish.

  • From Help Scout Docs: paste your Docs API key, from Your Profile, Authentication, API Keys. Private collections come over as team only.
  • From Zendesk: enter your subdomain. A public help centre needs nothing else. For restricted articles and drafts, add an admin's email and an API token.
  • From Intercom: paste an access token from an internal app that can read articles, and choose your region. Intercom doesn't say which articles are for which audience, so check any restricted ones before you publish.
  • From your docs site: crawl it on the Chat page first, then press Bring those pages in. They come over as plain text, so check headings and lists.
  • From files: upload Markdown, HTML or text files, or a zip of them, up to 9 MB and 500 articles. That covers an export from GitBook, Notion or Docusaurus, or a folder of your own. A folder becomes a collection, and title, description, tags and slug at the top of a file are read.

Credentials are used for that import only. They are never stored or shown again.

When it finishes, press Publish everything that was public there, or look through the drafts first. An article that was a draft or private in the old tool stays back, and so does anything your team wrote: you publish those yourself. Links between imported articles are pointed at their new addresses. Pictures stay where they were, so they go if you close the old help centre: check for them before you do.

Importing again adds new articles and refreshes ones nobody here has touched. An article your team has edited or published is left alone.

Write and publish

  1. On Articles, press New article.
  2. Give it a title and a one-line summary. Lists, search and agents read the summary first.
  3. Write it in Markdown: # headings, **bold**, - lists, links and tables.
  4. Choose a collection, and who can read it.
  5. Press Save and publish, or Save draft.

Everyone means your chat, your public pages and agents. Team only means your team, and agents with a token. The chat never reads a team-only article.

If two people edit the same article, the second to save is told, and keeps what they typed.

History and checking

Every change is a new version that says who made it: a person, an agent or an import. Open History on an article to see what changed in each version, and restore any of the last 50.

Press Mark as checked when you've confirmed an article is still right. Agents and readers see the date, so they can tell how far to trust it.

What agents can suggest

Review lists what agents have suggested: new articles, edits to published ones, and gaps they found. Each shows who suggested it, why, and what it's based on. An edit shows exactly what would change.

  • Accept an edit and it changes the live article straight away.
  • Accept a new article as a draft, or accept and publish it.
  • Dismiss anything you don't want.

If the article has changed since an agent suggested an edit, you're warned, and accepting it takes a second press.

Articles the AI drafts from gaps in your docs can join this list too: press Add to your knowledge base under a draft on Insights.

Try the agent

Under Try the agent, send a pretend request and watch the real agent handle it: your published articles, the same model, every safety check. You see its answer as that channel would show it, which articles and versions it read, or why it left the request for your team.

Choose where the request comes from: your website chat, email, WhatsApp, or a customer's AI assistant such as Muse or Instinct. Chat and email are live channels: on email the agent writes the reply and a person sends it. The others are simulated here: Swaru doesn't receive messages from them yet.

Run the sample set asks a mix across these channels: questions from your own articles, which the agent should answer, and requests only a person should take. It shows how many went the way they should.

Nothing is sent to anyone and no conversation is kept. It costs what an answer costs, from today's chat budget, and doesn't use up your plan's AI answers. You can send 40 requests and run the sample set 5 times a day, one at a time, and your company needs a plan.

Your public pages

An admin turns these on under Agents. You need a plan. Your help pages are then at getswaru.com/kb/your-company, with your published articles, search, and your brand colour.

Your chat answers from published articles whether these pages are on or off. With them on, its answers can link to the articles.

Connect an agent

With your public pages on, anything that speaks the Model Context Protocol (MCP) can search and read your published articles with no sign-in. Your server is at getswaru.com/kb/your-company/mcp. In Claude Code:

claude mcp add --transport http your-company https://getswaru.com/kb/your-company/mcp

Agents that don't use MCP can read the same articles from plain addresses:

  • /kb/your-company/llms.txt: what you have, with a link to each article.
  • /kb/your-company/llms-full.txt: every article in one file.
  • /kb/your-company/your-article.md: one article as Markdown. Asking for an article's page with Accept: text/markdown does the same.
  • /kb/your-company/search.json?q=refunds and /kb/your-company/articles.json.

Tokens

An admin makes tokens under Agents. Name each one after the agent that uses it: that name is what you see in Review and History. It's shown once, so copy it then. Send it in the Authorization header, as Bearer and the token.

  • Read only: also reads team-only articles, and works with your public pages off.
  • Read and suggest: also suggests new articles and edits, and reports gaps.

No token can publish or change an article. Revoke one under Agents and it stops working at once.

Limits

  • 2,000 articles and 100 MB of text with its history, each article up to 100,000 characters. Pictures show as links.
  • 200 suggestions waiting for review, and 50 a day from one token.
  • 20 tokens.
  • Each import: up to 5,000 articles and 2 hours. If it stops, start it again and it carries on.

Still stuck? Write to hello@getswaru.com.