> ## Documentation Index
> Fetch the complete documentation index at: https://docs.nogic.dev/llms.txt
> Use this file to discover all available pages before exploring further.

# Agent Canvas

> Ask AI on the canvas; answers come back as interactive diagrams

Nogic's Agent Canvas lets an AI client (Claude Code or Codex) answer questions by drawing diagrams on the visualizer canvas instead of generating prose. It's powered by a [local MCP server](/mcp/overview) that exposes canvas-rendering tools the AI can call.

<Note>
  **Works in any language.** The AI client reads your code with its own native `Read` / `Glob` / `Grep` tools, so `Cmd+K` works in Rust, Go, Java, Ruby, C++, or anything else, even though Nogic's [code graph](/features/visualizer) only parses JS / TS / Python today. The one exception is `render_code_tour` (high verbosity), which targets indexed symbols and so requires a parsed language.
</Note>

## The Director cursor

Press `Cmd+K` (or `Ctrl+K`) inside the visualizer. The on-canvas **Director cursor** opens an input bubble. Type a question, hit enter, and Nogic streams your prompt to the active CLI.

The cursor handles:

* Streaming narration from the AI as it works
* Auto-zooming to whatever the AI focuses on
* Walking through narrative steps one at a time (Next / Previous)
* Continuing the conversation: a second `Cmd+K` reuses the prior turn so you can drill in (`"Now show me the error path"`)

Conversations idle out after 30 minutes; after that, the next `Cmd+K` starts fresh.

## What the AI can render

The canvas supports several diagram families. The AI picks whichever fits the question best:

| Diagram                                    | Best for                                                            |
| ------------------------------------------ | ------------------------------------------------------------------- |
| **Generic tree** (`render` / `patch`)      | Architecture overviews, component hierarchies, free-form layouts    |
| **Dataflow** (`render_dataflow`)           | Pipelines, request flows, fan-out / fan-in patterns                 |
| **Sequence** (`render_sequence`)           | Request/response order, protocols, actor handshakes                 |
| **State machine** (`render_state_machine`) | Lifecycles, auth flows, retry/error transitions                     |
| **ER diagram** (`render_er_diagram`)       | Database schemas and table relationships                            |
| **Code tour** (`render_code_tour`)         | High-verbosity walks through real files / symbols in your workspace |

Each render comes with a step-by-step **narrative**. The cursor advances through steps, focusing the relevant primitives and reading the rationale. Patch tools (`patch`, `patch_dataflow`, `patch_sequence`, `patch_state_machine`) let the AI extend an existing diagram instead of redrawing it, useful for follow-up questions that drill into one piece.

See the [MCP overview](/mcp/overview) for the full tool surface.

## Verbosity & follow-ups

The cursor settings expose a verbosity dropdown:

* **Auto / low / medium**: concise diagrams, narrative-driven explanations
* **High**: switches to `render_code_tour`, which walks through real workspace symbols step by step with per-step rationale

You can also pick which CLI handles the turn (Claude Code vs Codex) and which model tier (`auto`, `cheap`, `medium`) directly from the cursor.

A second `Cmd+K` reuses the same conversation so follow-ups feel natural:

> "Walk me through how the auth flow works."
> *(diagram appears)*
> "Now show me what happens when the token is expired."
> *(diagram extends with the error path via patch ops)*

## Saving walkthroughs

Walkthroughs the AI generates are saved per workspace under `~/.nogic/workspaces/<hash>/saved_walkthroughs/`. They survive between sessions so you can revisit the answer without re-asking.

<Note>
  Saved walkthroughs are wiped once on extension upgrade if their payload shape changed. The on-canvas history is preserved, but new conversations start fresh.
</Note>

## Privacy considerations

Your prompts and the AI's tool calls go to whichever CLI you've configured (Claude Code or Codex). Nogic itself doesn't proxy or log them. The MCP server only exposes canvas-rendering tools; it never reads source files or sends them anywhere. Workspace files the AI reads come through the CLI's own native `Read` / `Glob` / `Grep` tools, governed by the CLI's permission model.

## Next

<CardGroup cols={2}>
  <Card title="Claude Code Setup" icon="terminal" href="/mcp/claude-code">
    Get `Cmd+K` working with Claude Code
  </Card>

  <Card title="Codex Setup" icon="robot" href="/mcp/codex">
    Get `Cmd+K` working with Codex
  </Card>
</CardGroup>