Connect Claude and your IDE over MCP

MITHUNAI runs a Model Context Protocol server, so AI clients such as Claude Code, Cursor and VS Code can search your knowledge and ask your assistants while they work. Answers come back grounded and cited, from the same answer engine as every other channel.

The MCP server is in beta: it offers three read-only tools and a feedback tool that rates an answer, and MITHUNAI does not yet connect to other MCP servers. Each text value in a tool result (a passage, a title or an answer) is limited to 512 characters and line breaks arrive as \n escapes, so a long answer is cut short. Use the HTTP API when you need the full answer text.

The server exposes three read-only tools and feedback, which records a rating of an answer you received and nothing else. None of them can change your knowledge, assistants or settings.

Tool	What it does	Arguments
`list_knowledge`	Lists the assistants you can ask and the knowledge bases you can search, with the IDs the other tools take. Archived assistants are not listed. Start here	None
`search_knowledge`	Returns the passages that best match a query, each with a citation. Does not generate an answer. Unless you name some, searches the collections your assistants answer from	`query` (required), `knowledge_base_ids`, `limit` (default 5, maximum 20)
`ask_assistant`	Asks an assistant and returns a grounded answer with citations, or an abstention. Each call is saved as a one-question conversation	`assistant_id`, `question` (both required)
`feedback`	Rates an answer `ask_assistant` returned as helpful or not, with an optional note. The rating appears with the console’s and the widget’s in analytics	`conversation_id`, `message_id`, `rating` (`helpful` or `not_helpful`), `note`

Connect a client

In the console, Integrations → MCP server shows the endpoint, the header to send and a request to try; there is nothing to switch on. You need:

The server URL: https://app.mithunai.com/mithunai/api/v1/mcp, or /mithunai/api/v1/mcp on your MITHUNAI host.
An API key. Create a dedicated key with the Member role for each person or tool. See API keys.

The server uses the streamable HTTP transport and authenticates with a bearer token.

Claude Code
Cursor
VS Code
Claude Desktop

claude mcp add --transport http mithunai https://app.mithunai.com/mithunai/api/v1/mcp \
  --header "Authorization: Bearer $MITHUNAI_API_KEY"

Add to ~/.cursor/mcp.json, or .cursor/mcp.json in a project:

{
  "mcpServers": {
    "mithunai": {
      "url": "https://app.mithunai.com/mithunai/api/v1/mcp",
      "headers": {
        "Authorization": "Bearer mithunai_sk_..."
      }
    }
  }
}

Add to .vscode/mcp.json:

{
  "inputs": [
    {
      "type": "promptString",
      "id": "mithunai-key",
      "description": "MITHUNAI API key",
      "password": true
    }
  ],
  "servers": {
    "mithunai": {
      "type": "http",
      "url": "https://app.mithunai.com/mithunai/api/v1/mcp",
      "headers": {
        "Authorization": "Bearer ${input:mithunai-key}"
      }
    }
  }
}

Claude Desktop connects to remote servers that need a custom header through the open-source mcp-remote bridge. Add to claude_desktop_config.json:

{
  "mcpServers": {
    "mithunai": {
      "command": "npx",
      "args": [
        "mcp-remote",
        "https://app.mithunai.com/mithunai/api/v1/mcp",
        "--header",
        "Authorization:${MITHUNAI_AUTH}"
      ],
      "env": {
        "MITHUNAI_AUTH": "Bearer mithunai_sk_..."
      }
    }
  }
}

A configuration file holding an API key is a secret. Keep it out of version control, or use your client’s secret or input mechanism as the VS Code example does.

Try it

Once connected, ask your client something that needs your documentation:

Example prompt for an MCP client connected to MITHUNAI

What the server does not do

One write, and only a rating. feedback stores a rating of an answer your own key received. No tool can add sources, change assistants or read settings, and an answer another key or person received cannot be rated: it is refused as not found.
No streaming. ask_assistant returns the complete answer in one response.
Temporary failures are tool errors. If the search index or model provider is unavailable, the tool returns an error result (isError: true) the client can show or retry. Permission and authentication problems are refused outright.
No browser access. MCP clients run on desktops and servers; the endpoint does not accept cross-origin browser requests.

Tool output is your ingested content. The server tells clients to treat it as content, not as instructions, and a well-behaved client will. The control that matters is that no tool can change anything.

Permissions and limits

The tools see exactly what the API key’s role can read, inside your organisation. An assistant in another organisation is refused as not found, and a search that names another organisation’s collection returns no passages, so neither reveals what exists elsewhere. MCP requests count against the API key’s normal budget and also against a separate MCP budget of the same size. Each API key has its own budgets, so give each MCP client its own key and an agent calling tools in a loop cannot exhaust another key’s budget. Clients that call from the same network address still share that address’s budget. See Rate limits. Each answer ask_assistant gives counts against your organisation’s monthly answer allowance, the same allowance the console, the widget and the API draw on. At the limit it is refused with 429 and the code quota_exceeded. search_knowledge and list_knowledge do not count. For the wire protocol, see the MCP endpoint reference.

MITHUNAI RAG

Knowledge sources

Answers and assistants

Deploy

Analyze

Connect Claude and your IDE over MCP

Connect a client

Try it

What the server does not do

Permissions and limits

​Connect a client

​Try it

​What the server does not do

​Permissions and limits

Connect a client

Try it

What the server does not do

Permissions and limits