Streaming End-to-End MCP Tool Use with an AI Model

See more AI Examples

Demonstrates Ai.UseMcp in a streaming Model Context Protocol (MCP) workflow. This is the same idea as the non-streaming MCP example, but the answer is streamed token-by-token and the automatic MCP tool call is driven by the application's event loop rather than happening invisibly inside Ask.

A tool-using conversation spans more than one streamed turn. In the first turn the model streams a js_function_call event; the example hands that event's JSON to StreamingJsToolCall, which routes MCP tools to the server and appends the result to the transcript, and the turn ends with null_terminator. The example then calls Ask again (with no new input, since the tool result is already in the transcript) so the model can continue and stream its final answer. The loop ends when a streamed turn finishes with no tool call outstanding.

Tip: Code to parse the returned JSON can be generated with Chilkat's online tool at https://tools.chilkat.io/jsonParse.

local chilkat = require("chilkat")

-- Streaming, end-to-end MCP example.  Same idea as the non-streaming MCP example, but the answer is
-- streamed token-by-token, and the automatic MCP tool call is driven by the application's event loop
-- instead of happening invisibly inside Ask.
--
-- A tool-using conversation spans more than one streamed turn:
--
--   turn 1:  the model streams a "js_function_call" event (it wants a tool).  We hand that event's
--            JSON to StreamingJsToolCall, which routes MCP tools to the server (Mcp.CallTool) and
--            appends the result to the transcript.  The turn then ends with "null_terminator".
--   turn 2:  we call Ask again (no new input -- the tool result is already in the transcript) so the
--            model can continue.  It streams the final answer and ends with "null_terminator", with
--            no pending tool call.
--
-- The loop ends when a streamed turn finishes with no tool call outstanding.
--
-- MCP server: DeepWiki -- a free, public, no-authentication Streamable HTTP server that answers
-- questions about public GitHub repositories.  Swap in any URL; set mcp.AuthToken before Connect if
-- it needs a bearer token.

-- 1) Connect to the MCP server.
local mcp = chilkat.Mcp.new()
-- mcp.AuthToken = "...";   // <-- only if the server requires a bearer token
local success = mcp:connect("https://mcp.deepwiki.com/mcp")
if not success then
    print(mcp.last_error_text)
    return
end

print("Connected to MCP server: " .. mcp.server_name .. " (version " .. mcp.server_version .. ")")

-- 2) Set up the AI conversation and register the MCP tools.
local ai = chilkat.Ai.new()
ai.provider = "anthropic"
-- The API key should come from a secure source rather than being hard-coded.
ai.api_key = "AI_PROVIDER_API_KEY"
ai.model = "claude-sonnet-5"

-- Automatic tool use requires a conversation (not a stateless query).
success = ai:new_convo("test_conversation", "You are a helpful assistant.  Use the available tools when they can help answer the question.", "")
if not success then
    print(ai.last_error_text)
    return
end

-- Namespace the server's tools with "deepwiki" -- e.g. "deepwiki_ask_question".  The Mcp object must
-- stay connected and in scope for the whole conversation.
success = ai:use_mcp(mcp, "deepwiki")
if not success then
    print(ai.last_error_text)
    return
end

-- Multiple MCP servers can be used in the same conversation.  Connect each server and call UseMcp
-- once per server, giving each a distinct prefix so their tool names cannot collide.  For example:
--     success = ai.UseMcp(mcpWeather,"weather");
--     success = ai.UseMcp(mcpGitHub,"github");

success = ai:input_add_text("Use the available tools to look up the GitHub repository \"modelcontextprotocol/modelcontextprotocol\", then give me a brief, sourced summary of which transports the Model Context Protocol defines.")
if not success then
    print(ai.last_error_text)
    return
end

-- 3) Start streaming.
ai.streaming = true
success = ai:ask("text")
if not success then
    print(ai.last_error_text)
    return
end

-- 4) Drive the streaming event loop.
local sbEventName = chilkat.StringBuilder.new()
local sbDelta = chilkat.StringBuilder.new()
local sbFullResponse = chilkat.StringBuilder.new()
local bConversationDone = false
local bToolCallPending = false
local bCaseSensitive = true
local bAbort = false
local loopGuard = 0
local result = 0

while (not bConversationDone) and (loopGuard < 8000) do
    loopGuard = loopGuard + 1
    result = ai:poll_ai(bAbort)
    if result < 0 then
        print(ai.last_error_text)
        return
    end

    if result == 0 then
        -- No event ready yet.
        ai:sleep_ms(100)
    else
        success = ai:next_ai_event(5000, sbEventName, sbDelta)
        if not success then
            print(ai.last_error_text)
            return
        end

        if sbEventName:contents_equal("js_function_call", bCaseSensitive) then
            -- (a) The model is requesting a tool call.  sbDelta holds the function-call JSON.
            -- StreamingJsToolCall runs the tool -- routing MCP tools to the server -- and appends the
            -- result to the transcript.  It must be called before this turn's null_terminator.
            print("[tool call requested]")
            print(sbDelta:get_as_string())
            success = ai:streaming_js_tool_call(sbDelta)
            if not success then
                print(ai.last_error_text)
                return
            end

            bToolCallPending = true
        elseif sbEventName:contents_equal("null_terminator", bCaseSensitive) then
            -- (c) This streamed turn has finished.
            if bToolCallPending then
                -- Continue the conversation so the model can use the tool result.  No new
                -- InputAddText -- the tool result is already in the transcript.
                bToolCallPending = false
                success = ai:ask("text")
                if not success then
                    print(ai.last_error_text)
                    return
                end
            else
                bConversationDone = true
            end
        else
            -- (b) An "empty" event reports nothing this poll.  (d) Any other event is a normal
            -- streamed text delta, which we display and accumulate.
            if not sbEventName:contents_equal("empty", bCaseSensitive) then
                print("Event: " .. sbEventName:get_as_string() .. "  Delta: " .. sbDelta:get_as_string())
                sbFullResponse:append_sb(sbDelta)
            end
        end
    end
end

print("Final streamed response:")
print(sbFullResponse:get_as_string())
print("----")

-- (Optional) The full transcript, including the tool call and its result.
local convoJson = chilkat.JsonObject.new()
convoJson.emit_compact = false
success = ai:export_convo("test_conversation", convoJson)
if not success then
    print(ai.last_error_text)
    return
end

print("Full Conversation:")
print(convoJson:emit())

-- 5) Close the MCP session.
success = mcp:close()
if not success then
    print(mcp.last_error_text)
    return
end