GraphAI - MCP(Model Context Protocol) Agnet

GraphAI now supports MCP (Model Context Protocol).

There are two ways to invoke an MCP server: via network or through inter-process communication. In this implementation, we focus only on inter-process communication using the command configuration.

While MCP provides various methods like init and resources, this implementation only makes use of tools/list and tools/call. These two methods are sufficient to interact with most MCP servers.

Unlike other GraphAI agents, MCP requires an initial connection to the server before it can be used. This involves executing a command on the host machine to start the server. Therefore, the MCP initialization process must occur before GraphAI is initialized.

GraphAI example is here.

https://github.com/receptron/graphai-agents/blob/main/protocol/mcp-agent/tests/run_graph.ts

npm is

https://www.npmjs.com/package/@graphai/mcp_agent

1. Configuring each service via mcpConfig

Each service is defined as a pair of command and args. This pattern can be extended to other applications as well.

The mcpConfig uses unique service names (in this example, filesystem and filesystem2) as keys, and specifies configuration values as values. The following example demonstrates this using the server-filesystem MCP server, which operates on the local file system.

The filesystem configuration shows how to run the npm package directly via npx, while filesystem2 specifies the executable file after installing the npm package. Both methods function the same way, and typically, you only need to specify one of them. However, both are shown here as examples of different configuration approaches.

export const mcpConfig = {
  filesystem: {
    command: "npx",
    args: ["-y", "@modelcontextprotocol/server-filesystem", path],
  },
  filesystem2: {
    command: __dirname + "/../../../node_modules/@modelcontextprotocol/server-filesystem/dist/index.js",
    args: [path],
  },
};

2. Initializing MCP and connecting to the server

MCP must be started before running GraphAI. Since the startup time may vary, it’s recommended to include a waiting period to ensure the server is ready. If GraphAI is run before MCP finishes initializing, the list of tools might be returned as empty.

The function that initializes MCP returns mcpClients. Since mcpClients is required when calling MCP, it is passed as a config to the GraphAI instance.
It is also needed when closing the connection to the MCP server.

  const mcpClients = await mcpInit(mcpConfig)
  await setTimeout(2000);

3. Running GraphAI

When creating a GraphAI instance, mcpClients is passed via the config.
Since MCP is divided into multiple agents, mcpClients is provided globally (although it's also possible to pass it individually).

After defining the graph data and registering the necessary agents, GraphAI can be executed.

The mcpToolsListAgent returns a merged list of tools from all services defined in the configuration. Tool names are prefixed with the service name to handle namespace collisions. The tools are provided in both raw format and OpenAI-compatible format.

The mcpToolsCallAgent executes the selected tool via tools/call, using the namespaced tool name as-is.

  const graphData = {
    version: 0.5,
    nodes: {
      list: {
        agent: "mcpToolsListAgent",
        isResult: true,
        console: true,
      },
      llm: {
        agent: "openAIAgent",
        inputs: {
          tools: ":list.llmTools",
          prompt: "get all file list on " + path
        },
      },
      call: {
        console: {before: true, after: true},
        agent: "mcpToolsCallAgent",
        inputs: {
          tools: ":llm.tool"
        },
      },
    },
  };
  const graphai = new GraphAI(graphData, {...vanilla,  mcpToolsListAgent, mcpToolsCallAgent, openAIAgent }, { config: { global: { mcpClients } } });
  const result = await graphai.run();
  

Result is

{
  "response": {
    "content": [
      {
        "type": "text",
        "text": "[FILE] 1.txt\n[FILE] 2.txt"
      }
    ]
  }
}

4. Disconnecting from the MCP server

Be sure to explicitly disconnect from the MCP server when your batch job or server (e.g., Express) shuts down.
mcpClients is also passed when disconnecting.

  await setTimeout(500);
  mcpClose(mcpClients);