MCP Client

Connect external MCP servers to Vaquill and use their tools directly in chat

Vaquill’s MCP Client lets you connect any Model Context Protocol server - Notion, GitHub, Slack, databases, or your own custom server - so its tools become available inside Vaquill’s AI chat alongside the built-in legal tools.

This page covers connecting external MCP servers to Vaquill. If you want to connect Vaquill’s tools to your AI editor (Claude, Cursor, etc.), see Vaquill MCP.

Built-in Legal MCP Servers

Two legal research servers are available to all users by default - no setup required:

CourtListener

US federal and state court opinions, dockets, PACER data, and citations. 19 tools. No auth needed.

CanLII

Canadian federal and provincial court decisions and legislation. 7 tools. Bilingual (EN/FR).

These system servers can be toggled on/off individually in Settings > MCP Servers.

Connect Your Own MCP Servers

1. Go to Settings

   Open Settings > MCP Servers in the dashboard.

2. Add a server

   Click Add Server and provide:

   • Name - a label for the server (e.g., “Notion”)
   • URL - the MCP endpoint URL
   • Transport - streamable_http (recommended) or sse
   • Headers (optional) - authorization headers (encrypted at rest)
3. Sync tools

   Click Sync to discover the server’s available tools. You can preview each tool’s description and parameters.

4. Use in chat

   Enable Agent Mode in any chat. The server’s tools automatically appear in the AI’s tool list and can be called autonomously.

Agent Mode must be enabled for MCP tools to work in chat. The AI calls MCP tools as part of its autonomous tool-calling loop.

How It Works

When you chat with Agent Mode enabled:

1. Tools from all your enabled MCP servers are discovered and cached
2. A semantic relevance filter prunes irrelevant tools based on your query (so the AI only sees tools that matter)
3. The AI decides whether to call an MCP tool, just like it decides to search documents or fetch a URL
4. Tool results are returned to the AI with safety wrappers to prevent prompt injection
5. Tools are prefixed with the server name (e.g., notion__create_page) to avoid collisions

Security

• SSRF protection - server URLs are validated against private IP ranges, loopback, and reserved addresses
• Encrypted headers - authorization headers are AES-256-GCM encrypted before storage and never returned in API responses
• Prompt injection mitigation - tool results are wrapped with explicit “untrusted data” markers
• No tool shadowing - MCP tools cannot override built-in tools (web search, corpus search, etc.)
• Rate limiting - 100 tool executions per hour, 5 sync requests per minute

Limits

Limit	Value
Max servers per user	10
Max tools per server	10
Max total MCP tools	15
Discovery timeout	10 seconds
Execution timeout	30 seconds
Tool result size cap	8,000 characters

Transport Protocols

Protocol	Best For	Overhead
Streamable HTTP (recommended)	Most servers, stateless	Near-zero
SSE (legacy)	Servers that only support SSE	~300 KB per call

Related

• Agent Mode - Required for MCP tools to work in chat
• Vaquill MCP - Connect Vaquill’s tools to your AI editor
• CourtListener MCP - US court data server
• CanLII MCP - Canadian legal data server