> ## Documentation Index
> Fetch the complete documentation index at: https://docs.enconvo.ai/llms.txt
> Use this file to discover all available pages before exploring further.

# Knowledge Base MCP

> Let Claude, Cursor, Codex, and other MCP clients on your Mac search and read the knowledge bases you share.

Enconvo can serve your knowledge bases to MCP-compatible AI clients on the same Mac. Clients can search and read the knowledge bases you share, and nothing else; access is read-only. This is separate from **Settings → MCP Servers**, which adds external tools to Enconvo's Agents.

## Connect a client

1. Open the [Knowledge Base](https://docs.enconvo.ai/ai/knowledge-base.md) and click the gear (**Knowledge Base Settings**), then **AI Clients (MCP)**.
2. Turn on **Enable Knowledge Base MCP**. It's off by default, and while it's off, connected clients are refused.
3. Choose what clients can see. **Share this knowledge base with AI clients** covers the one you have open, and **Shared knowledge bases** lists them all. Dictation History and Memory are hidden until you turn them on.
4. Under **Connect a client**, pick your client (Claude Code, Cursor, Codex, or Claude Desktop) and copy its setup.
5. Ask the client to list knowledge bases, then search one and read a matching item.

The local endpoint is:

```text theme={null}
http://localhost:54535/knowledge_base/mcp
```

It uses the Streamable HTTP transport. Keep Enconvo running while the client uses it. Claude Desktop connects through a small bridge, `mcp-remote`, so its setup needs Node.js; follow the client-specific instructions shown in the dialog.

## Available tools

| Tool | Purpose |
| - | - |
| `kb_list` | List visible knowledge bases |
| `kb_search` | Search a visible knowledge base |
| `item_list` | List items in a visible knowledge base |
| `item_read` | Read indexed item text in pages |

The tools are read-only. This endpoint doesn't expose general Memory or Recordings tools. Reading an item can return its source path or URL as metadata.

## Access and processing

The server accepts loopback connections only and doesn't require a token. Turning it on lets local processes read the knowledge bases you share. A connected AI client can send what it gets back to its own model provider, so review that client's settings as well.

Search uses the embedding model set under **Search** in **Settings → Memory**. When it runs on Enconvo Cloud, the search text is sent there. "Local endpoint" doesn't mean every part of retrieval and answering runs offline.

## Troubleshooting

| Problem | What to do |
| - | - |
| Connection refused | Make sure Enconvo and its local API are running. |
| Forbidden / 403 | Check that **Enable Knowledge Base MCP** is on, that the client connects from this Mac, and that its configuration matches the dialog. |
| Knowledge base not found | Turn the knowledge base on under **Shared knowledge bases**. Hidden and missing knowledge bases look the same to clients. |
| Opening the URL in a browser shows 405 | A browser GET isn't an MCP request. Connect with an MCP client. |
| A long document comes back incomplete | Ask the client to request further pages with `item_read`. |

To stop all access, turn off **Enable Knowledge Base MCP**. To narrow what clients can retrieve, turn off individual knowledge bases under **Shared knowledge bases**.

## Related

> [card] [Knowledge Base](https://docs.enconvo.ai/ai/knowledge-base.md)
>
> Build the knowledge bases clients will search.

> [card] [MCP Servers](https://docs.enconvo.ai/features/mcp.md)
>
> Add external tools to Enconvo's own Agents.

> [card] [CLI and Local API](https://docs.enconvo.ai/advanced/cli-local-api.md)
>
> Other ways to reach Enconvo from your Mac.


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.