As MCP Tools — AI Functions

Automatic MCP Registration

When an agent is initialised, the AIFunctionToolRegistrar queries all active AI Functions for that agent and tenant, then registers each one as an MCPTool:

public class AIFunctionToolRegistrar
{
    public async Task RegisterForAgentAsync(
        Guid agentId, Guid tenantId,
        MCPToolRegistry registry,
        CancellationToken ct)
    {
        // Load all active AI Functions for this agent + shared tenant functions
        var functions = await _store.GetActiveForAgentAsync(agentId, tenantId, ct);

        foreach (var fn in functions.Where(f => f.InputSchema != null))
        {
            registry.Register(new MCPTool
            {
                Schema = new ToolDefinition
                {
                    Name        = fn.Name,
                    Description = fn.Description ?? fn.Name,
                    InputSchema = JsonDocument.Parse(fn.InputSchema!)
                },
                Handler = async (input, ctx, ct2) =>
                {
                    var result = await _runtime.ExecuteAsync(fn, input, new AIFunctionContext
                    {
                        AgentId  = ctx.AgentId,
                        UserId   = ctx.UserId,
                        TenantId = ctx.TenantId
                    }, ct2);

                    return JsonSerializer.Serialize(result);
                }
            });
        }
    }
}

Creating an AI Function MCP Tool — Step by Step

Write the Function Code Write a JavaScript execute(input) function that performs the desired logic.

Define Name and Description Give the function a unique name (snake_case) and a tool description that tells the LLM when to call it.

Define Input Schema Write the JSON Schema for the function's input parameters. This is what the LLM uses to construct tool call arguments.

Assign to Agent In the agents-app, assign the AI Function to the target agent (or leave AgentId null for tenant-wide availability).

Test via Chat Open a chat with the agent and trigger the function. Use the Context Inspector to verify the tool was called and the result injected.

AI Function vs. C# Tool: When to Use Each

Scenario	Use AI Function	Use C# Tool
Data formatting / calculation	Yes — pure logic, no I/O needed	Overkill
Database query	No — no DB access in sandbox	Yes
External API call	No — no network in sandbox	Yes
Logic that admins may want to change	Yes — no deployment needed	No — requires developer
Complex multi-service orchestration	No — too complex for sandbox	Yes

← Inputs and Outputs Next: Flow Studio Node →