> The full index is at https://docs.takepilot.ai/llms.txt

# Make videos with the plugin

![From asking to receiving: 1. Ask (what, for whom), 2. Estimate (you say OK), 3. Make (voice, image, clip), 4. Check (preview), 5. Receive (My files)](https://docs.takepilot.ai/images/diagrams/en/plugin-loop.svg)

With the TakePilot plugin, you can ask the AI agent you already use for "a 15-second video of this offering". The agent writes the plan and shows an estimate; when you say OK, TakePilot's servers make the voice, images and clips and put them into one video.

## Connect {#setup}

![The web's "Connect an AI agent" screen: a tab per agent, with copyable commands](https://docs.takepilot.ai/images/screens/en/agent-connect.webp)

First create an account on the web ([app.takepilot.ai](https://app.takepilot.ai)) and add credits. Then open "Settings" → "Connect an AI agent" on the web and pick your agent's tab; the commands can be copied from there. The first time a tool is used, a browser opens: sign in to TakePilot and allow it.

| Agent | What you add |
|---|---|
| Claude Code | The plugin (2 commands) |
| Claude Desktop | The plugin zip (chat mode) |
| ChatGPT desktop app | A custom MCP server (no skills or commands; it reads how to make the video from `get_playbook`) |
| Codex CLI | The plugin (3 commands) |
| Grok Build | The plugin (1 command) |
| Other AI agents | The MCP server and the skill zip ([steps](#other-agents)) |

> [!NOTE]
> To make videos in an organization's workspace, an owner or admin of the organization first turns on "Make from AI agents" ([Use AI agents in an organization](https://docs.takepilot.ai/en/workspaces-and-organizations#use-agents-in-an-organization)). You always connect with your own account.

### The MCP and the skill {#mcp-and-skill}

Making videos from an agent takes two things: the MCP and the skill. The plugin gives you both.

| | What it is | Without it |
|---|---|---|
| MCP | The tools: make voice, images and clips, put them into one video, save it | The agent can't call TakePilot |
| Skill (`takepilot-ad`) | How to make the video: the brief, hook and appeal options, the NG-expression check, the estimate, checking the result | The agent can call the tools but doesn't have the steps |
| Plugin | The MCP, the skill and the commands, as one set | |

If you only added the MCP URL, you are halfway there. [Add the skill too](#other-agents).

### Claude Code {#claude-code}

```
/plugin marketplace add TakePilot/takepilot-plugin
/plugin install takepilot@takepilot
```

Open `/mcp`, choose `takepilot` and authenticate; press "Allow" when the browser opens. Then ask, for example, `/takepilot:make a 30-second TikTok Shop video for a toner`.

### Claude Desktop {#claude-app}

1. On the web's "Connect an AI agent" page, open the Claude Desktop tab and download the plugin zip. Do not unzip it.
2. In Claude Desktop's chat mode, open **Customize** → **Plugins** → **Add** (top right) → **Upload plugin**.

   ![The Add menu with Upload plugin marked](https://docs.takepilot.ai/images/plugin/claude-add-menu.webp)

3. Choose the zip and press **Upload**.

   ![The Upload a plugin screen with the Upload button marked](https://docs.takepilot.ai/images/plugin/claude-upload.webp)

4. On the plugin's **Connectors** tab, open the `takepilot` row.

   ![The plugin's Connectors tab with the takepilot row marked](https://docs.takepilot.ai/images/plugin/claude-connector.webp)

5. Press **Connect**.

   ![The Connect button under "You're not connected to takepilot yet."](https://docs.takepilot.ai/images/plugin/claude-connect.webp)

6. **Add custom connector** opens with the name and URL filled in. Press **Continue**.

   ![The bottom of the Add custom connector dialog with Continue marked](https://docs.takepilot.ai/images/plugin/claude-add-connector.webp)

7. Keep **Sign in now** and **Register automatically** selected and press **Add**.

   ![The authentication choices with Sign in now and Register automatically marked](https://docs.takepilot.ai/images/plugin/claude-auth-options.webp)

8. When the row shows **Connect** again, press it. In the browser press **Continue connecting** and allow TakePilot. **Connected** means you are done.

   ![The Connectors tab with Connected on the takepilot row](https://docs.takepilot.ai/images/plugin/claude-connected.webp)

9. In a new chat, ask "Make a 30-second TikTok Shop video for a toner with takepilot".

- On Team and Enterprise, the organization's Owner adds the connector; then each person connects with their own account.
- Chat has nowhere to keep files, so finished videos come as download URLs.
- Pressing Connect in the Code tab can show "Couldn't start sign-in … different server URL". Press it in chat mode instead.

#### Fewer permission prompts {#claude-app-permissions}

In Claude Desktop's chat, every call to a TakePilot tool asks for permission. Set the read-only tools to always allowed and you will be asked far less often.

1. In Claude Desktop, open **Customize** → **Connectors** and choose `takepilot`.
2. Under **Tool permissions**, set **Read-only tools** to **Always allow**.
3. Leave **Write/delete tools** on **Needs approval**.

Read-only tools only look at your offerings, My Files and balance. They change nothing and spend no credits.

Keep the tools that spend credits (making voices, images and clips, building, rendering) on **Needs approval**. Before making anything, the agent shows the estimate and asks you to choose, in a multiple-choice question, whether to go ahead. Each call then asks for your permission, so credits never go without you knowing.

- The setting stays with the connector; you do not need to set it again in a new chat.
- On Team and Enterprise, limits set by the organization's Owner take precedence.
- The menu names follow Claude's [connector help article](https://support.claude.com/en/articles/11176164).

### ChatGPT desktop app {#chatgpt}

ChatGPT takes no plugin: connect it as a custom MCP server. You get TakePilot's tools and the guide for asking (`get_playbook`). Use **Chat** mode; **Work** mode shows no tools.

1. Copy the MCP server URL `https://api.takepilot.ai/v1/mcp`.
2. Open **Plugins** in the sidebar → **Add** (top right) → **Add custom MCP server**.

   ![The Add menu with Add custom MCP server marked](https://docs.takepilot.ai/images/plugin/chatgpt-add-menu.webp)

3. Enter `takepilot` as **Name** and the URL as **Server URL**. Tick **I understand and want to continue** and press **Create as a plugin**.
4. Press **Connect**, then sign in to TakePilot in the browser and allow it.
5. Choose **Chat** at the top, pick `@takepilot` in the input box and ask "Call get_account and tell me my balance". If the balance comes back, it works.

   ![Chat mode asking @takepilot for get_account](https://docs.takepilot.ai/images/plugin/chatgpt-chat.webp)

ChatGPT can't take the skill, so when you ask for a video, add "read get_playbook first". If the tools look out of date, press **Manage** → **Refresh tools**.

### Codex CLI {#codex}

```
codex plugin marketplace add TakePilot/takepilot-plugin
codex plugin add takepilot@takepilot
codex mcp login takepilot
```

The Codex desktop app is not supported. On a shared computer, run `codex mcp logout takepilot` when you are done.

### Grok Build {#grok-build}

```
grok plugin install TakePilot/takepilot-plugin
```

Trust the plugin to load it. Then open `/mcps`, choose `takepilot` and sign in.

### Other AI agents {#other-agents}

Agents that cannot take a plugin get the MCP and the skill separately. The MCP alone only brings the tools, so carry on through step 2.

**1. Connect the MCP**

Add it as a remote MCP server.

| Item | Value |
|---|---|
| MCP server URL | `https://api.takepilot.ai/v1/mcp` |
| Transport | HTTP (streamable HTTP) |
| Sign-in | OAuth (sign in to TakePilot in the browser and allow) |

**2. Add the skill**

On the web's "[Connect an AI agent](https://app.takepilot.ai/settings/agent)" page, open the "Others" tab and download the skill zip.

| Your agent | How to add it |
|---|---|
| Has a screen for adding skills | Upload the zip as is |
| Reads skills from a folder | Unzip it and put `takepilot-ad` in that folder (see your agent's docs for where) |
| Can't take skills | Add "read get_playbook first" when you ask for a video |

**3. Ask whether the skill is there**

Ask your agent "can you use the takepilot-ad skill?". "Connected agents" on the web only shows the MCP connection, not whether the skill was added.

The skill changes as TakePilot updates. If you added it by hand, it doesn't update itself, so download and add it again from the same tab now and then.

> [!NOTE]
> On claude.ai on the web, skills are added with **Add** under Settings > Customize > Skills. We checked that `takepilot-skill.zip` uploads and that the skill loads and plans a video (2026-11-06); we haven't made a whole video this way yet. claude.ai can't reach your computer, so the MP4 is written on the server, and it asks about the cost first. Anthropic's guide says skills need code execution turned on.

### Check the connection {#check-the-connection}

Under "Connected agents" on the web's "Connect an AI agent" page, you see each agent's name and when it was last used. "Disconnect" stops that agent within 10 minutes.

## How to ask {#how-to-ask}

Include these five and the first try gets close.

| Include | Example |
|---|---|
| Offering | The name of a registered offering |
| Who it is for | People in their 20s, after work |
| The first 1–3 seconds | The sound of opening the box, close-up |
| One message | Not too sweet |
| How many versions | 3 hooks × 2 appeals |

> A 15-second TikTok video of Forest Scent Eau de Parfum for women in their 20s, with Sakura talking. 3 different hooks

The web's "Prompt builder" collects what you pressed "Use" on (offerings, avatars, templates and more) and turns it into text to paste into your agent. To hand over just one thing, press "Prompt" on the page of a video, offering or other item: text pointing to it is copied, ready to paste.

![The Prompt builder screen: what you press "Use" on in Offerings, Templates, Stock and My files gathers here](https://docs.takepilot.ai/images/screens/en/picked.webp)

> [!TIP]
> Make one video first to settle the flow, then add versions with other hooks or appeals. It keeps costs down.

## What the agent does {#what-it-does}

- **Registers offerings**: give it a URL or photos and it creates the offering, with its kind and CVs ([Kind and CV](https://docs.takepilot.ai/en/avatars-and-products#kind-and-cv)).
- **Estimates**: shows the cost and asks you before making anything. An estimate over the per-call limit is not made ([Plugin costs and limits](https://docs.takepilot.ai/en/plugin-costs)).
- **Makes**: asks TakePilot for voice, images and clips, and assembles them in a template.
- **Checks**: looks at stills and a short preview for text over faces or the offering.
- **Makes versions**: remakes only what changes and reuses shared clips.
- **Fixes**: reads review comments and fixes only the parts named, as a new version ([Ask someone to review](https://docs.takepilot.ai/en/client-review)).
- **Delivers**: rendered videos land in My files ([Find videos in My files](https://docs.takepilot.ai/en/my-files)).

## Video models {#video-models}

Unless you name one, the agent picks for you: scenes MiniMax H3 Flash can make are made with Flash, and the rest (some scenes with people, for example) with MiniMax H3. Other models (Seedance, Kling, Luma, Higgsfield) are used only when you name them in your request; the agent tells you the price difference in one line.

## Only on the web {#web-only}

| Task | Where |
|---|---|
| Publish to Community | "Community" |
| Buy credits; raise a cap | "Credits"; "Settings" → "Caps" |

> [!WARNING]
> Make avatars only from faces and voices you have the right to use. Do not make videos that imitate celebrities or claim effects for real brands.

## Next steps {#next-steps}

- [Plugin costs and limits](https://docs.takepilot.ai/en/plugin-costs)
- [Register avatars and offerings](https://docs.takepilot.ai/en/avatars-and-products)