Skip to main content

MCP Servers

MCP (Model Context Protocol) is an open standard that allows AI Agents to communicate with external tool servers. Through MCP, you can connect capabilities such as filesystem access, web search, database queries, and code repository browsing to Elftia in a standardized way, enabling Agents to invoke these external tools directly during a conversation.

Why MCP

An Agent's built-in capabilities are limited. With MCP, you can:

  • Expand the toolset — Register the capabilities of any external service as tools that the Agent can call
  • Standardize the interface — Tools from different sources communicate using a unified protocol, eliminating the need to write custom adapters for each one
  • Connect on demand — Enable only the servers needed for the current task, keeping the tool list from becoming unwieldy
  • Reuse across projects — The same set of MCP server configurations can be shared across multiple Agents

Three Transport Protocols

Elftia supports three MCP transport types. Choose the one that fits your use case:

Transport typeHow it connectsUse caseExample
StdioLaunches a local child process, communicates via standard input/outputLocal tools, CLI wrappers, npm/pip packagesnpx @modelcontextprotocol/server-filesystem
SSELong connection via HTTP Server-Sent EventsRemote services, scenarios requiring real-time pushhttps://api.example.com/mcp/sse
HTTPRequest/response based on Streamable HTTPRemote REST-style MCP serviceshttps://api.example.com/mcp

:::tip How to Choose

  • Choose Stdio for most cases — MCP server packages installed via npm or pip typically use Stdio
  • Choose SSE or HTTP for remotely hosted services — such as cloud-hosted MCP services provided by third parties
  • SSE vs. HTTP — SSE is based on the older MCP spec; HTTP is based on the MCP 2025-03-26 spec (Streamable HTTP). Prefer HTTP for new services :::

Tool Discovery and Usage Flow

Once you configure an MCP server, Elftia automatically handles the following:

MCP server starts / connects

Elftia automatically discovers the list of tools the server provides

Tools are registered in the Agent's available toolset

The Agent calls tools as needed during a conversation

The tool returns results; the Agent integrates them into its reply

Tool naming format: Each MCP tool is registered as mcp__<server-name>__<tool-name>. For example, the read_file tool provided by a server named filesystem is identified in the Agent as mcp__filesystem__read_file.

MCP and Agents

MCP servers can be paired with Agents flexibly:

  • Automatic mode (default) — Tools from all enabled MCP servers are automatically available to all Agents
  • Manual mode — Select which MCP servers to use before each conversation
  • Agent binding — Associate specific MCP servers with specific Agents for dedicated tool configurations

Feature Summary

FeatureDescription
Multiple transport protocolsStdio, SSE, and HTTP connection options
Automatic tool discoveryAfter connecting, the tool list is fetched automatically and cached for 5 minutes
JSON bulk importAdd multiple servers at once via JSON configuration
Dependency auto-detectionAutomatically checks whether node/npm/uv/uvx and other CLI tools are installed
Connection testingOne-click test of server connection status and available tools
Tool disablingSelectively disable specific tools from a server
Trust managementDistinguish between trusted and untrusted servers
Official presetsBuilt-in one-click installation configurations for common MCP servers

Next Steps