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

# Connect your agent

> Set up VibeFlow in Claude, Claude Code, Cursor, Codex, OpenClaw, or Hermes. One paste, OAuth sign-in, no keys.

Every setup below ends the same way: your agent gets the full canvas toolset — list and create canvases, author frames, screenshot them, follow your design systems, export decks. Pick your client and go.

<Note>
  **OAuth is the default.** You don't create or paste an API key. The first time your agent touches a VibeFlow tool, your browser opens, you sign in, click **Allow access**, and you're done. Prefer a long-lived key for scripts or CI? See [Canvas MCP](/mcp/overview#api-key-auth).
</Note>

***

## The one-paste setup

The fastest path for agents that can configure their own MCP servers (OpenClaw, Hermes, Claude Code, and most agentic CLIs). Paste this into the chat:

```
Set up VibeFlow from vibeflow.ai/claude
```

Swap the URL for your agent's page — [`vibeflow.ai/claude`](https://vibeflow.ai/claude), [`vibeflow.ai/codex`](https://vibeflow.ai/codex), [`vibeflow.ai/claw`](https://vibeflow.ai/claw), [`vibeflow.ai/cursor`](https://vibeflow.ai/cursor), [`vibeflow.ai/hermes`](https://vibeflow.ai/hermes). Each page serves agent-readable setup instructions, so the agent wires itself up and asks you to approve access once.

If your client can't self-configure, use the per-client steps below.

***

## Claude (web & desktop)

<Steps>
  <Step title="Open Connectors">
    In Claude, go to **Settings → Connectors**.
  </Step>

  <Step title="Add a custom connector">
    Click **Add custom connector**, name it `VibeFlow`, and paste:

    ```
    https://tool.vibeflow.ai/mcp
    ```
  </Step>

  <Step title="Approve access">
    Claude opens VibeFlow in your browser. Sign in and click **Allow access**. The canvas tools appear in Claude's tool list immediately.
  </Step>
</Steps>

***

## Claude Code

```bash theme={null}
claude mcp add --transport http vibeflow https://tool.vibeflow.ai/mcp
```

Then type `/mcp` inside Claude Code and pick **vibeflow** to sign in — it opens your browser for the one-time approval. No `--header` flag needed; OAuth handles auth.

***

## Cursor

**One click:** hit **Install in Cursor** on [vibeflow.ai/cursor](https://vibeflow.ai/cursor). Cursor prompts you to sign in on first use.

**Or manually** — add to `~/.cursor/mcp.json`:

```json theme={null}
{
  "mcpServers": {
    "vibeflow": {
      "url": "https://tool.vibeflow.ai/mcp"
    }
  }
}
```

Verify under **Cursor Settings → Tools & MCP** — you should see the VibeFlow tools listed with a green dot after the OAuth round-trip.

***

## Codex

Add to `~/.codex/config.toml`:

```toml theme={null}
[mcp_servers.vibeflow]
url = "https://tool.vibeflow.ai/mcp"
```

Or in the UI: **Settings → MCP Servers → Streamable HTTP**, name `vibeflow`, URL `https://tool.vibeflow.ai/mcp`.

***

## OpenClaw

Paste into your OpenClaw chat:

```
Set up VibeFlow from vibeflow.ai/claw
```

Or add to your MCP configuration:

```json theme={null}
{
  "mcpServers": {
    "vibeflow": {
      "url": "https://tool.vibeflow.ai/mcp"
    }
  }
}
```

***

## Hermes Agent

Paste into your Hermes chat:

```
Set up VibeFlow from vibeflow.ai/hermes
```

Or add the same `mcpServers` block as OpenClaw above to Hermes' MCP configuration.

***

## Copilot (VS Code)

Create `.vscode/mcp.json` in your workspace:

```json theme={null}
{
  "servers": {
    "vibeflow": {
      "type": "http",
      "url": "https://tool.vibeflow.ai/mcp"
    }
  }
}
```

***

## Anything else

Any MCP client that speaks Streamable HTTP with OAuth works:

| Field     | Value                                       |
| --------- | ------------------------------------------- |
| MCP URL   | `https://tool.vibeflow.ai/mcp`              |
| Transport | Streamable HTTP                             |
| Auth      | OAuth 2.1 (automatic — leave headers empty) |

***

## Prompts to try

Once connected, verify the loop end-to-end:

```
Create a canvas called "MCP test" and design a single 1080×1350
Instagram post announcing a summer sale. Screenshot it and show me.
```

Then push it harder:

* *"List my canvases and screenshot the four most recent frames on the latest one."*
* *"Read the design system on my 'Brand refresh' canvas, then add a pricing section frame that matches it."*
* *"Design a three-slide launch deck for our v2 release — title, value prop, pricing — then export the canvas as PPTX and give me the download link."*
* *"Duplicate the hero frame twice and give me a darker variant and a warmer variant, side by side."*
* *"Take the copy in this doc and turn it into a five-frame LinkedIn carousel."*

***

## Getting good designs out of your agent

The VibeFlow design agent follows a house style guide; your agent can too.

<Tip>
  Tell your agent: *"Load the VibeFlow canvas guide and design guide before you start."* That's two tool calls — `get_canvas_guide` (workflow + frame-HTML rules) and `get_design_guide` (visual hierarchy, spacing, typography, color discipline, and the anti-"AI-generated look" bans). Agents that skip these produce noticeably more generic work.
</Tip>

Three habits that separate good sessions from bad ones:

1. **Make it look.** After any visual change, have the agent call `screenshot_canvas_frames` and fix what's off. Agents that judge layouts from HTML alone ship broken spacing.
2. **Attach a design system.** Frames authored against a [design system](/concepts/design-systems) inherit your palette, type, and components. Without one, every agent invents its own brand.
3. **Iterate surgically.** For small changes, point the agent at `update_canvas_frame` (string replace) instead of full rewrites — faster, and it can't mangle the rest of the frame.

***

## Troubleshooting

* **Tools don't appear after setup** — complete the OAuth round-trip first: trigger any VibeFlow tool once, approve in the browser, then retry. Some clients only list tools after the first successful auth.
* **Browser never opens** — your client may not support OAuth discovery. Fall back to an [API key](/mcp/overview#api-key-auth) with an `Authorization: Bearer` header.
* **Agent behaves oddly mid-session** — long agent sessions drift. Restart the session; the connector and its auth survive.
* **"Insufficient credits" errors** — tool calls bill like in-app usage. See [Credits](/reference/credits).

## Next

<CardGroup cols={2}>
  <Card title="Tool reference" icon="list-check" href="/mcp/tools">
    Every canvas tool, signature, and behavior.
  </Card>

  <Card title="Canvas MCP overview" icon="plug" href="/mcp/overview">
    Transport, auth modes, and pricing.
  </Card>
</CardGroup>
