Generative AI — User Guide

Overview

Chat with an AI coding agent wired to InfiniWorkflow's own MCP server, so it can edit the live workflow graph, generate and compile C++ plugins, drive AI Studio, inspect state, and control playback, all from natural-language requests. Both Claude (Anthropic, via the claude/Claude Code CLI) and ChatGPT (beta) (OpenAI, via the codex CLI) are supported - pick between them with the chatbox's "CLI" dropdown. This page covers the in-app chatbox specifically; see the full MCP / AI Agent page for the underlying tool list and for connecting an external agent (Claude Desktop, AnythingLLM, ...) instead.

1. Prerequisites

The fastmcp Python package - installed automatically the first time you use the chatbox (or the MCP page's "Launch" button), or manually via python -m pip install fastmcp. See app/packages/mcp.json.
The CLI for whichever agent you want to use: claude (Claude Code) for Claude, or codex (OpenAI's Codex CLI) for ChatGPT. Sign in once in a terminal so it has its own authentication configured.

No Google Cloud / Vertex AI setup, API keys, or region selection are needed - the CLI handles its own authentication, the same way it would if you ran it in a terminal yourself.

2. Opening the Chatbox

Click the Generative AI icon in the application toolbar. The first time you use it on a machine, it asks which CLI command to run (defaults to claude) before opening the chat - after that, the same setting is available in the dialog's own "Command" field, alongside a "CLI" dropdown for choosing between Claude and ChatGPT. Both are shared, via the same cookie, with the MCP page's own "Launch an agent" command field.

Example: the Generative AI chatbox, mid-conversation

You

Distribute the AI Welding branch to my GPU server and show me the FPS after

Claude

Found a contiguous 3-node branch (Canny → Find Contours → Draw Contours). Distributing it to 192.168.1.42 now - Profile shows 18fps → 54fps after the move.

Generate a Python sample app for this workflow

Add a Telemetry output for the frame rate

Enter a prompt here

3. A Single Conversation, Not Separate Modes

There's no "Workflow" vs. "Plugin" mode anymore - one free-form conversation decides on its own which tools a message calls for (build/edit a graph, generate a plugin, run AI Studio, inspect state, control playback, and more), the same way an external Claude Code session connected via MCP would. A fresh conversation shows a row of example prompts spanning that range - click one to drop it into the input (see the two chips in the mock above).

4. Real Conversation History

Each conversation is a genuine, persistent multi-turn session, not a one-shot request:

Feature	Mechanism
Threads	The sidebar (toggled via `#history-toggle`) lists past conversations by title and last-updated time - click one to load its full transcript, or the icon to start a new one.
Memory	Each thread keeps its own persistent CLI session, so later messages in the same conversation have full context of earlier ones - reattached via the CLI's own resume mechanism even after an InfiniWorkflow restart.
Deletion	Hovering a thread row shows an to delete it, closing its session and dropping its saved transcript.
Storage	`app/chatbox_threads.json` on disk, not cookies.

5. What Happens When You Send a Message

The agent calls MCP tools directly against the live, running workflow - graph edits appear on the canvas as they happen (over the same live-update mechanism an external Claude Code/Desktop session would use), so there's no separate "review and import" step for a generated graph.
If the agent drafts a new C++ plugin, it opens the same "Generated Plugin Files" viewer as before so you can read the code and click Download or Compile yourself - see Plugin Generation on the MCP page for the full flow.
The agent's own reply text appears in the chat pane underneath your message.