Graph Copilot

Graph Copilot is an agent workbench plugin for the editor. It builds and edits the graph by conversation: describe what you want in plain language and the AI adds nodes, connects ports, and sets parameters, applying each batch of operations as one undo step. It can also run the current graph after you approve, run bounded experiment studies on in-memory copies of the graph, and keep a browsable conversation history.

:::note Availability Graph Copilot is built on two CodefyUI features: the plugin frontend extension API and the unified LLM proxy endpoint (/api/llm/chat). Both are in CodefyUI 1.3.0 and later. If cdui --version reports an older version, run cdui update before installing. :::

Installation

cdui plugin install graph-copilot

You can also open the Plugin Center in the editor and install graph-copilot from the catalog. Installation through the Plugin Center loads the panel immediately. After installing from a terminal, reload the editor by pressing F5. The Graph Copilot panel appears as a floating widget in the editor.

Plugin source and issues: github.com/CodefyUI/CodefyUI-Plugin-Graph-Copilot

Quick start

1. Install the plugin (above).
2. Click the Graph Copilot button (a chat-bubble icon) in the bottom-right corner of the editor to open the panel.
3. Click Settings (gear icon), choose a provider, and paste your API key. To use OpenAI Codex (ChatGPT), click Sign in and approve access in the tab that opens. Select a model; use Refresh to fetch the provider's model list.
4. Type a request such as Build a small MLP classifier and press Enter.
5. The AI streams its plan while adding and connecting nodes on the canvas. Press Ctrl+Z to undo its last batch of operations, or send another message to refine it.

The browser stores the provider and key, so you only need to configure them once. The following sections describe each feature.

Panel views

The panel has four views, chosen with the buttons at its top:

• Chat: the conversation with the agent.
• Lab: experiment studies run from Chat, and studies imported from a file (Import portable study). A study runs each variant on an in-memory copy of the current graph and ranks the results. The canvas changes only when you ask the agent to apply the winner and the winner changes only parameters: it is then applied as one undo step, unless the graph changed while the study ran.
• History: earlier conversations. Click one to resume it, or start a new one with + New chat.
• Settings: provider, model and reasoning effort, API key or sign-in, and run notifications.

Choosing an LLM provider

Click the Settings (gear) icon in the Graph Copilot panel to configure the provider and key.

Provider	Notes
OpenAI API	Standard https://api.openai.com/v1 endpoint. Requires an OpenAI API key. Billed per token.
OpenAI Codex (ChatGPT)	Uses OAuth with your ChatGPT account through the Codex CLI PKCE flow and client ID. It uses ChatGPT subscription quota rather than API credits and remains subject to ChatGPT usage limits and OpenAI's terms. Tokens are stored on the server in llm/codex_auth.json under the user-data directory. Everyone who uses that server shares one signed-in account, and Sign out clears the tokens for everyone. The OAuth callback listens on localhost:1455 (or 1457) in the server process. Complete sign-in within 5 minutes in a browser on the machine that runs the server.
OpenRouter	Aggregates many providers under one key. The proxy sends requests to https://openrouter.ai/api/v1; select your preferred model.
Claude API	Anthropic's API, accessed through CodefyUI's proxy. The proxy translates the OpenAI-compatible request format. Requires an Anthropic API key.
Custom (OpenAI-compatible)	Any server that implements the OpenAI /chat/completions endpoint. For example, you can use a local Ollama instance at http://localhost:11434/v1. Set the base URL and, if required, a key.

The proxy also provides POST /api/llm/models, which lists a provider's models for Refresh. POST /api/llm/codex/login, GET /api/llm/codex/status, and POST /api/llm/codex/logout support ChatGPT sign-in. The same controls are available under Settings → LLM Providers. Only OpenAI API and OpenAI Codex (ChatGPT) use reasoning_effort. The value must be 1–64 characters of lowercase letters, digits, _ or -, starting with a letter; any other value is 422 for every provider. The proxy refuses ultra for OpenAI Codex (ChatGPT) (400) and forwards other values unchanged. The editor supplies the session token required by the POST routes.

Key handling

API keys are stored in localStorage under a namespace private to Graph Copilot. Each request sends the selected key to the local CodefyUI backend. /api/llm/chat forwards the key and messages to the configured provider and streams the response back. It does not log or persist the key or messages. Each provider has fixed upstream hosts; only Custom uses a base URL that you supply. These keys are separate from the environment variables CODEFYUI_OPENAI_API_KEY / OPENAI_API_KEY and CODEFYUI_ANTHROPIC_API_KEY / ANTHROPIC_API_KEY, which only the LLMChat node reads; the proxy reads none of them.

Usage

Sending a request

Type your request in the chat input and press Enter (or click Send). Examples:

• "Add a two-layer MLP with ReLU activations"
• "Connect the CrossEntropy node to the output of the last Linear"
• "Set the hidden size on Linear-1 to 512"

The AI returns a plan and then applies operations. Each tool call appears as a step under the reply, for example Edit graph with add_node ×2, connect ×1, and shows whether it succeeded.

Conversation history

Conversations are stored in the browser (Graph Copilot's namespaced localStorage), newest first, at most 50; past that the oldest is dropped. They are not tied to a graph. Opening the editor starts a new conversation; open History to resume an earlier one.

Stop and retry

Click Stop during a stream to cancel the in-flight request; text that already streamed stays in the conversation. When a request fails, the panel shows Request failed with a Retry button that resends your last message and its attachments.

Undoing AI edits

Each batch of operations the AI applies is one undo snapshot. Press Ctrl+Z (or Cmd+Z on macOS) once to undo the last batch. When the AI corrects itself within one reply, it applies more than one batch, and each needs its own Ctrl+Z.

Server-side studies

The Lab runs a study in the page that started it, so closing that tab stops the study. CodefyUI also offers server-side studies that keep running after the tab closes, survive a restart, enforce budgets and are listed in the Studies panel. A plugin delegates its studies to them through the plugin jobs routes (POST /api/plugins/graph-copilot/jobs with kind: "study", listed in the API Reference). The plugin's chat keeps planning the experiment and asking for your approval; the server runs the approved plan. Moving the Lab onto these routes is a change in the plugin's own repository, and until a plugin release does so the Lab runs in the page as described above.

Tips

• Give context about what you are building: "I am building a vision classifier with a ResNet backbone" helps the AI make better choices.
• If the AI adds a node type that does not exist in your palette, it will be skipped and reported — use cdui plugin install to add the required pack first.
• Graph Copilot reads the current graph state and the full node palette before each request, so it knows what types are available and what is already on the canvas.

See also

• Graph Copilot documentation — experiments, attachments, graph runs, and the agent contract.
• Plugin Frontend Extensions — the JS API that Graph Copilot is built on.
• Plugins — the plugin pack system.
• API Reference — the /api/llm/chat streaming endpoint.