Automatically discover tools for Azure OpenAI Realtime API

September 24, 2025 9 min read BEN ABT

Azure now provides a unified Realtime API for low‑latency, multimodal conversations over WebRTC or WebSockets. If you’ve used the earlier preview versions (for example the GPT‑4o realtime preview), the new generation model is simply called gpt-realtime and the API follows the same event-driven pattern: you open a session, configure defaults via session.update, stream input and receive streaming output (text, function calls, audio, etc.).

In this post I’ll focus on one topic that’s easy to get wrong but crucial for production: how to register tools and let the model discover and use them dynamically at runtime.

Tooling

The Realtime API supports tools (function calling) just like the Responses API. Tools are declared on the session using a session.update event with a tools array of function descriptors. From there, the model can propose function calls during a response and you return the tool results back into the conversation to let the model continue.

Many apps don’t have a fixed set of functions; available tools can depend on tenant, plan, user role, feature flags or backend availability. Hardcoding every tool in the session is brittle and wastes tokens. The better approach is a tiny “tool registry” you always expose, which allows the model to discover the current catalog and execute tools on demand.

This article shows a minimal, robust pattern:

Register only three stable functions at session start: list_tools, describe_tool and run_tool.
Seed the instructions so the model knows to discover before calling anything.
Implement a small dispatcher on your server that resolves and validates tool calls.

The result: dynamic tools without re‑opening sessions or wasting tokens on large static schemas.

Session bootstrap: register a tool registry

Send a session.update as the first message after opening the WebSocket/WebRTC session. Keep parameters concise to reduce token usage.

JSON
 1{
"type": "session.update",
"session": {
  "instructions": "You are a realtime assistant. Before attempting external actions, discover capabilities via list_tools/describe_tool. Use run_tool only with validated arguments. If a tool is missing, ask the user.",
  "turn_detection": { "type": "server_vad" },
  "tools": [
    {
      "type": "function",
      "name": "list_tools",
      "description": "Return the list of currently available tools. Include name, short summary and whether arguments are required.",
      "parameters": {
        "type": "object",
        "properties": {},
        "additionalProperties": false
      }
    },
    {
      "type": "function",
      "name": "describe_tool",
      "description": "Return the JSON schema for a tool by name (arguments, required, constraints).",
      "parameters": {
        "type": "object",
        "properties": {
          "name": { "type": "string", "description": "Tool name" }
        },
        "required": ["name"],
        "additionalProperties": false
      }
    },
    {
      "type": "function",
      "name": "run_tool",
      "description": "Execute a tool by name with JSON arguments and return a structured result.",
      "parameters": {
        "type": "object",
        "properties": {
          "name": { "type": "string" },
          "arguments": { "type": "object" }
        },
        "required": ["name", "arguments"],
        "additionalProperties": false
      }
    }
  ]
}
46}

With this registry in place, your instructions can nudge the model to:

Call list_tools when it suspects an external action is needed.
Call describe_tool to fetch the exact argument schema for a chosen capability.
Call run_tool once it has a valid argument object.

Event loop: detect tool calls and return results

When the model wants to use a tool, it emits tool‑call events as part of the streaming response. Your server should:

Accumulate the streaming arguments for each tool call (IDs allow parallel calls).
Validate and execute the tool on your side.
Send a tool result back into the conversation referencing the call ID, then continue the response.

A compact TypeScript sketch for a WebSocket session (server‑to‑server) might look like this:

TypeScript
 1// Simplified sketch: handle tool calls over Realtime WebSocket
 2import WebSocket from "ws";
 3
 4type ToolCall = { id: string; name: string; args: string };
 5
 6const ws = new WebSocket("wss://<your-aoai-endpoint>/openai/realtime?api-version=<latest>&deployment=<gpt-realtime-deployment>", {
 7  headers: { Authorization: `Bearer ${process.env.AZURE_OPENAI_TOKEN}` }
 8});
 9
10// Collect partial tool-call arguments by id
11const pending: Record<string, ToolCall> = {};
12
13ws.on("open", () => {
14  // 1) Send session.update with registry (see JSON above)
15  ws.send(JSON.stringify(/* ...session.update payload... */));
16});
17
18ws.on("message", async (raw) => {
19  const evt = JSON.parse(raw.toString());
20
21  // Tool call argument streaming (delta events)
22  if (evt.type === "response.output_tool_call.arguments.delta") {
23    const id = evt.item?.id as string;
24    const name = evt.item?.name as string;
25    pending[id] ??= { id, name, args: "" };
26    pending[id].args += evt.delta ?? "";
27    return;
28  }
29
30  // Tool call completed (we have the full JSON args string)
31  if (evt.type === "response.output_tool_call.completed") {
32    const id = evt.item?.id as string;
33    const call = pending[id];
34    if (!call) return;
35    const args = JSON.parse(call.args || "{}");
36
37    // 2) Dispatch locally
38    const result = await dispatch(call.name, args);
39
40    // 3) Return tool result back into the conversation
41    ws.send(JSON.stringify({
42      type: "conversation.item.create",
43      item: {
44        type: "tool_result",
45        tool_call_id: id,
46        content: [
47          { type: "output_text", text: JSON.stringify(result) }
48        ]
49      }
50    }));
51  }
52});
53
54async function dispatch(name: string, args: any) {
55  switch (name) {
56    case "list_tools": return listTools();
57    case "describe_tool": return describeTool(args.name);
58    case "run_tool": return runTool(args.name, args.arguments);
59    default: return { error: `Unknown tool: ${name}` };
60  }
61}
62
63function listTools() {
64  return [
65    { name: "weather.getByCity", summary: "Get current weather by city", argsRequired: true },
66    { name: "calendar.search", summary: "Find events", argsRequired: false }
67  ];
68}
69
70function describeTool(name: string) {
71  if (name === "weather.getByCity") {
72    return {
73      name,
74      parameters: {
75        type: "object",
76        properties: { city: { type: "string" }, unit: { type: "string", enum: ["c","f"] } },
77        required: ["city"],
78        additionalProperties: false
79      }
80    };
81  }
82  return { name, parameters: { type: "object", properties: {} } };
83}
84
85async function runTool(name: string, args: any) {
86  if (name === "weather.getByCity") {
87    // call your backend / cache / API
88    return { city: args.city, unit: args.unit ?? "c", temp: 22.3 };
89  }
90  return { error: `Tool not implemented: ${name}` };
91}

Notes:

The Realtime API streams tool arguments in deltas; concatenate the delta chunks until you receive a completion event for that tool call.
Use the tool_call_id you got from the model when sending back the tool_result so the model can continue the same response coherently.
The exact event names may evolve; always cross‑check with the latest Azure documentation.

C#: minimal WebSocket skeleton

If you prefer .NET, a thin ClientWebSocket loop works fine. You still exchange JSON events just like above.

C#
 1using System;
 2using System.Collections.Generic;
 3using System.Net.WebSockets;
 4using System.Text;
 5using System.Text.Json;
 6using System.Text.Json.Serialization;
 7using System.Threading;
 8using System.Threading.Tasks;
 9
10namespace RealtimeSample
11{
internal static class Program
{
  private static async Task Main()
  {
    Uri uri = new Uri("wss://<your-aoai-endpoint>/openai/realtime?api-version=<latest>&deployment=<gpt-realtime-deployment>");
    using ClientWebSocket ws = new ClientWebSocket();
    ws.Options.SetRequestHeader("Authorization", $"Bearer {Environment.GetEnvironmentVariable("AZURE_OPENAI_TOKEN")}");
    await ws.ConnectAsync(uri, CancellationToken.None);
20
    // Send session.update with your registry
    byte[] sessionUpdate = Encoding.UTF8.GetBytes(JsonSerializer.Serialize(new
    {
      type = "session.update",
      session = new
      {
        instructions = "Discover tools first, then call run_tool with validated args.",
        tools = new object[]
        {
          new { type = "function", name = "list_tools", parameters = new { type = "object", properties = new { } } },
          new { type = "function", name = "describe_tool", parameters = new { type = "object", properties = new { name = new { type = "string" } }, required = new[] { "name" } } },
          new { type = "function", name = "run_tool", parameters = new { type = "object", properties = new { name = new { type = "string" }, arguments = new { type = "object" } }, required = new[] { "name", "arguments" } } }
        }
      }
    }));
    await ws.SendAsync(new ArraySegment<byte>(sessionUpdate), WebSocketMessageType.Text, true, CancellationToken.None);
37
    // Read loop (simplified)
    byte[] buffer = new byte[64 * 1024];
    Dictionary<string, StringBuilder> pending = new Dictionary<string, StringBuilder>();
41
    while (ws.State == WebSocketState.Open)
    {
      WebSocketReceiveResult result = await ws.ReceiveAsync(new ArraySegment<byte>(buffer), CancellationToken.None);
      string json = Encoding.UTF8.GetString(buffer, 0, result.Count);
      JsonElement evt = JsonDocument.Parse(json).RootElement;
47
      string? type = evt.GetProperty("type").GetString();
      if (type == "response.output_tool_call.arguments.delta")
      {
        string id = evt.GetProperty("item").GetProperty("id").GetString()!;
        string delta = evt.GetProperty("delta").GetString() ?? string.Empty;
        if (!pending.TryGetValue(id, out StringBuilder? sb))
        {
          sb = new StringBuilder();
          pending[id] = sb;
        }
        sb.Append(delta);
      }
      else if (type == "response.output_tool_call.completed")
      {
        JsonElement item = evt.GetProperty("item");
        string id = item.GetProperty("id").GetString()!;
        string name = item.GetProperty("name").GetString()!;
        StringBuilder? argsSb;
        string argsJson = pending.TryGetValue(id, out argsSb) ? argsSb.ToString() : "{}";
        object? args = JsonSerializer.Deserialize<object>(argsJson);
68
        object resultObj = await DispatchAsync(name, args);
        byte[] toolResult = JsonSerializer.SerializeToUtf8Bytes(new
        {
          type = "conversation.item.create",
          item = new
          {
            type = "tool_result",
            tool_call_id = id,
            content = new object[] { new { type = "output_text", text = JsonSerializer.Serialize(resultObj) } }
          }
        });
        await ws.SendAsync(new ArraySegment<byte>(toolResult), WebSocketMessageType.Text, true, CancellationToken.None);
      }
    }
  }
84
  private static Task<object> DispatchAsync(string name, object? args)
  {
    return Task.FromResult<object>(new { ok = true, name, args });
  }
}
90}

Why this pattern works

Dynamic by design: your actual tool set can change per tenant/user/session. The registry maps the current world.
Token‑efficient: only three small functions live in the system prompt. Detailed schemas are fetched on demand.
Safer: you keep a single validator/dispatcher (run_tool) with allow‑lists, quotas and telemetry.
Parallel‑ready: the model can propose multiple tool calls; you correlate by tool_call_id and execute concurrently.

Best practices

Keep instructions short and explicit about discovery order: list → describe → run.
Validate everything in run_tool: name allow‑list, argument schema validation, timeouts and retries.
Limit payload sizes. If a tool returns large data, return a short summary plus a handle/URL.
Instrument: log call IDs, latency, failures and user/session correlation IDs for support.
Security: never allow arbitrary code execution; don’t expose internal service names or secrets in tool errors.
Versioning: include a version field in describe_tool responses so the model can re‑query when versions change.

Common pitfalls

Registering dozens of functions directly in tools bloats tokens and slows first response.
Missing tool_call_id when sending the result back → the model won’t connect the dots and may stall.
Returning free‑form text for complex data → prefer compact JSON (stringified if needed) so the model can reason over it.
Not handling partial argument deltas → you’ll parse broken JSON. Always accumulate until completion.

References

Azure OpenAI Realtime – WebSockets quickstart and event reference
Function calling concepts and parallel tool calls

With a small registry and a disciplined event loop, you can keep sessions lean, enable truly dynamic capabilities and let gpt‑realtime do what it does best: orchestrate.

Automatically discover tools for Azure OpenAI Realtime API

Tooling

Session bootstrap: register a tool registry

Event loop: detect tool calls and return results

C#: minimal WebSocket skeleton

Why this pattern works

Best practices

Common pitfalls

References

Automatic TTL in Azure Cosmos DB with .NET

Azure IoT: send messages with C#

Production Rate Limiting in ASP.NET Core

BEN ABT

Let's Work Together

New Platforms

Modernization

Training & Consulting

Automatically discover tools for Azure OpenAI Realtime API

Tooling

Session bootstrap: register a tool registry

Event loop: detect tool calls and return results

C#: minimal WebSocket skeleton

Why this pattern works

Best practices

Common pitfalls

References

Related articles

Let's Work Together

New Platforms

Modernization

Training & Consulting