Company Knowledge Base for Claude and ChatGPT — How It Works in Practice
Your team uses Claude or ChatGPT every day. And every day, someone pastes a wall of company documentation into the chat window to get a useful answer — the HR policy they can't find, the sales playbook section relevant to a specific deal, the engineering onboarding steps for a new contractor. It works, but it's friction: find the doc, copy the text, paste it in, hope it fits in the context window. There's a cleaner setup. This post walks through connecting your company knowledge base to Claude and ChatGPT via MCP, so the AI already has access to the right content without the paste-and-pray workflow.
What this looks like in practice
Once your company knowledge base is connected as an MCP endpoint:
- An engineer asks Claude: "What's the process for getting contractor access to the staging environment?" Claude pulls the answer from your engineering wiki, cites the exact page, and returns the steps. No pasting the wiki page first.
- A sales rep asks ChatGPT: "What objection-handling language do we use for the 'too expensive' pushback?" ChatGPT queries the indexed sales playbook and returns the relevant section.
- An HR manager asks Claude: "Summarize the parental leave policy for employees in Singapore." Claude finds the relevant policy pages in the indexed HR docs and returns a cited summary.
The AI isn't guessing from general knowledge. It's querying a search index built from your actual company docs, and every answer it returns links back to the source page so anyone can verify and read the full context.
What to index
Any knowledge that lives as a public or accessible URL can be indexed. Common starting points:
- HR and People Ops. Employee handbook, benefits documentation, leave policies, onboarding checklists, offboarding steps. Questions that employees ask HR repeatedly — and that HR would rather not answer via Slack message for the fiftieth time — are ideal.
- Sales playbooks and enablement. Objection handling, competitive positioning, pricing breakdowns, customer success case studies, ICP profiles. Reps asking "what do we say about Competitor X?" get cited, consistent answers instead of asking a colleague.
- Engineering wikis. Architecture decisions, runbook procedures, deployment checklists, environment setup guides, API authentication flows. New hires and contractors can query the wiki through Claude without digging through Confluence or Notion.
- Product documentation. If your product docs are publicly accessible, index them. Support agents using Claude to answer tickets get accurate, current answers from the same docs your customers see.
- Legal and compliance templates. Standard contract terms, vendor review checklists, data processing agreement summaries. Index the approved reference docs so the AI returns the company-blessed answers rather than general legal information.
Public vs private. WebToMCP indexes publicly accessible URLs. If your knowledge base lives behind an auth wall (Confluence, internal Notion, password-protected Gitbook), you'll need to either export it to a public-accessible location or use a self-hosted indexing setup. We're working on auth-gated indexing — drop us a note if that's your situation.
How to set it up
The setup is three steps regardless of which AI client your team uses:
- Index your knowledge base. Sign in to webtomcp.net with Google, click New knowledge base, and submit the root URL of the site or section you want indexed. If your company has separate sites for HR docs, engineering wiki, and product docs, create a separate knowledge base for each — they'll each get their own endpoint URL. Indexing takes 1–10 minutes for most sites.
-
Copy the MCP endpoint URL. When the knowledge base flips to Ready, copy the endpoint URL from the dashboard. It looks like
https://mcp.webtomcp.net/batches/<id>/mcp?key=wmcp_…. Share this URL with the team members who need it. -
Connect each person's AI client.
- Claude Desktop: Settings → Connectors → Add → paste the endpoint URL → restart Claude. Full walkthrough →
- ChatGPT: Settings → Connectors → Add custom connector → paste the endpoint URL → Add. Full walkthrough →
- Cursor or other MCP clients: Settings → MCP → add server with the endpoint URL.
The free tier covers a single knowledge base with ~500 AI questions per month — enough to evaluate the workflow. Hobby at $9/month covers five knowledge bases and 5,000 questions. For a team with a handful of knowledge bases (HR, sales, engineering), that's the starting point.
Keeping the knowledge base current
WebToMCP re-crawls connected sites weekly by default. If your HR policy changes or you publish a new sales playbook section, the index will pick it up on the next scheduled crawl. For time-sensitive updates — a new competitor tear-down that reps need today, a policy change that takes effect immediately — click Re-crawl in the dashboard for an instant refresh.
The re-crawl is incremental: we detect pages that haven't changed (via ETag / content hash) and skip re-embedding them. A weekly refresh on a 200-page docs site typically takes a few minutes and doesn't burn significant query quota.
What this isn't
A couple of clarifications that come up often:
- It's not a chatbot widget on your website. The MCP endpoint is used by your team's AI clients — Claude Desktop, ChatGPT, Cursor. It doesn't embed a chat bubble on your company's website. (If you want a public-facing chatbot on a site, that's a different product category.)
- It's not replacing your knowledge base platform. You still maintain your docs in Notion, Confluence, or wherever they live. WebToMCP indexes the public-facing version of those docs and makes them queryable through AI. You're not migrating content; you're adding a query layer on top.
- It's not training the AI on your data. The AI doesn't get retrained with your company's content. It retrieves from your index at query time. Your content stays in the index, not in the model weights. Privacy policy →
Where to start
Pick the highest-friction knowledge base in your company — the one where people most often paste wall-of-text docs into Claude or ChatGPT to get an answer. Index that one first. Run it for a week. See whether the AI's answers improve in accuracy and citation quality compared to what your team was getting from paste-and-pray.
If it works — and for most content-heavy knowledge bases it will — expand to the next one. The free tier costs nothing to test with. Sign in with Google to get started.
Related reads
- Your entire website as a live AI knowledge base — how WebToMCP works end-to-end.
- Connect your website to Claude Desktop using MCP
- Connect your website to ChatGPT (every tier, including Free)
- llms.txt vs MCP — when to use a static file vs a live query server
Questions? developer@webtomcp.net. Or sign in with Google to index your own site (free).