To connect an AI agent to your help centre with MCP, add your help centre’s remote MCP server address to the agent (in Claude under Connectors, in Claude Code with claude mcp add, in Cursor and VS Code in an mcp.json file) and sign in when asked. The agent can then search and read your articles and, if you allow it, draft new ones and updates for your team to approve.
The examples use AskBase’s server, https://askbase.com/mcp, because it’s a real address you can try. The steps are the same for any help centre or documentation tool that offers a remote MCP server.
What MCP is
The Model Context Protocol (MCP) is an open standard for connecting AI applications to other systems. A server describes a set of tools (“search articles”, “propose an update”), and any client that speaks MCP, such as Claude, Cursor or VS Code, can discover and call them. You connect once, and the agent decides when to use each tool while it works on what you asked.
For a help centre, that means an agent that can look things up in your articles while it’s writing code, answering a question or reading a pull request, and can write documentation changes straight into your review queue instead of into a chat window you then copy from.
Remote and local servers
MCP servers run in one of two ways, and the difference decides how you connect.
- Local servers run as a program on your own computer, started by the client, and talk to it over standard input and output (stdio). They’re common for tools that work on local files.
- Remote servers run on the provider’s side at a web address, using the protocol’s Streamable HTTP transport. There’s nothing to install: you give the client the address.
Signing in with OAuth
Remote servers usually need to know who you are. The MCP specification uses OAuth 2.1 for this: when you add the server, the client opens the provider’s sign-in page, you approve what the agent may do, and the client keeps a token. You never paste a password or an API key into the agent, and you can revoke access from the provider’s side.
On AskBase the approval screen asks for two permissions: reading your help centres, and proposing changes for review. Untick the second and the agent can only read.
Connect your agent
Claude (web and desktop)
Go to Customize → Connectors, choose + Add, then Add custom connector. Enter a name and the server address, continue, and sign in when asked. Custom connectors work on Free, Pro, Max, Team and Enterprise plans; Free is limited to one.
On Team and Enterprise, an owner first adds the connector for the organisation under Organization settings → Connectors (Add, then Custom, then Web). Members then find it under Customize → Connectors and choose Connect.
https://askbase.com/mcp
Claude Code
Add the server from your terminal:
claude mcp add --transport http askbase https://askbase.com/mcp
Then run /mcp inside Claude Code and choose the server to sign in. By default the server is added for the current project only. Add --scope user to have it in every project, or --scope project to share it with your team through a .mcp.json file in the repository.
Cursor
Add the server to ~/.cursor/mcp.json for every project, or .cursor/mcp.json in one project, then sign in from Cursor’s MCP settings:
{
"mcpServers": {
"askbase": {
"url": "https://askbase.com/mcp"
}
}
}VS Code
Add it to .vscode/mcp.json in a workspace, or to the mcp.json in your user profile. VS Code handles the sign-in:
{
"servers": {
"askbase": {
"type": "http",
"url": "https://askbase.com/mcp"
}
}
}ChatGPT
Open Plugins in ChatGPT, choose the plus button, then Add custom MCP server. Enter a name and the full address including /mcp, set up authentication, and create it. Whether you can add custom servers depends on your plan and your workspace’s policies, so check OpenAI’s current help if the option isn’t there.
Other clients
Any client that supports remote MCP servers over Streamable HTTP with OAuth can connect the same way: give it the address, and it finds where to sign in from the server.
What the agent can do
A server lists its tools, and the client shows the agent each one’s name and description. These are the tools AskBase’s server offers today. Tools marked proposes need the permission to propose changes.
| Tool | What it does | Access |
|---|---|---|
list_ | The help centres you can reach, in every workspace you're in: id, name, workspace, your role there, and the public address. | reads |
search_ | Full-text search over a help centre's published articles, with highlighted snippets. | reads |
list_ | Every article in a help centre: id, title, path, status and when it was last updated. | reads |
read_ | One article in full, as Markdown. | reads |
list_ | Screens AskBase captured by signing in to the product: name, purpose, controls, how to reach each one, and a screenshot URL that stays current, to embed in articles with Markdown: . | reads |
list_ | Questions customers asked in the last 30 days that the help centre couldn't answer, grouped and most asked first: good things to write about. | reads |
list_ | Changes waiting for review (or approved, or rejected), including ones you proposed: so you don't propose the same thing twice. | reads |
read_ | One suggested change in full: the proposed text, the article it would replace, and its checklist. | reads |
propose_ | Put a new article you wrote into the review queue. | proposes |
propose_ | Put a rewrite of an existing article into the review queue. | proposes |
request_ | Have AskBase itself plan and draft up to six new articles or updates from a request in words ('document the form builder'), using the help centre, the company website and product screens. | proposes |
check_ | How recent request_articles runs are going: queued, running, done or failed, what was planned, and how many drafts are waiting for review. | reads |
dismiss_ | Mark a content gap as not worth an article (spam, a one-off, handled elsewhere). | proposes |
upload_ | Store a screenshot or image (PNG, JPEG, GIF or WebP, up to 8 MB, base64) and get back a URL to embed in a proposed article with Markdown: . | proposes |
Every tool runs as you, with your role in the workspace that owns the help centre. Viewers can read but not propose. Contributors and members can propose. No tool publishes, approves or deletes anything, whatever your role. Proposals show in the review queue as coming from the app, such as “Claude (via MCP)”, so reviewers know where they came from.
Prompts that work
Agents do best with a concrete job and the context to do it. A few that work well against a help centre:
After a release
We just shipped the changes in this pull request. Find the help articles that describe anything it changed, and propose updates to each one. Put anything you couldn't confirm from the code under "To check".
Filling gaps
List the content gaps for our help centre. For the three most asked, check whether an article already covers them, then propose a new article for each one that isn't covered.
Answering from the docs while you work
A customer asks whether exports include archived items. Search our help centre and tell me what it says, with the article links. If it doesn't say, propose a short article once I've confirmed the answer.
A tidy-up
Read every article in the Billing category. List any that mention a plan name or price that differs from the pricing article, and propose corrections.
On AskBase, an agent can also hand a whole request to AskBase itself with request_articles (“document the new form builder”), which plans and drafts up to six articles or updates in the background.
Keeping a person in charge
An agent that can write to your help centre is useful because it’s fast, and risky for the same reason. Whatever tool you use, check three things before you connect one:
- Changes wait for approval. A draft in a review queue costs nothing if it’s wrong. A published article that’s wrong is answered to customers.
- You can limit it to reading. Many uses, such as looking things up while coding, need no write access at all.
- Changes are labelled with where they came from, so reviewers can be more careful with machine-written drafts.
The API and the agent skill
MCP suits agents a person is working with. For scripts and CI, such as proposing doc updates on every merge, a REST API with a key fits better. AskBase keys are made by an admin with one of three levels of access: Read only, Propose changes, or Full access. A key is never an admin, and replies it writes wait for a person by default. The API and MCP documentation lists every operation.
An agent skill is a set of instructions that teaches an agent how to do a job well. AskBase’s agent skill covers keeping articles current after you ship, filling the gaps customers ask about and drafting inbox answers, over either MCP or the API. For Claude Code, save it as a personal skill:
mkdir -p ~/.claude/skills/askbase curl -o ~/.claude/skills/askbase/SKILL.md https://askbase.com/skill.md