Generate images, video and audio in coding agents

A coding agent's main model must be a text model: it has to chat and call tools. Image, video and audio models only produce content, so they cannot be the main model; picking one in the model selector fails.

Instead, let the text model call the media models. Two ways:

Both use your own API key and are billed exactly like the API: images per image, video per second, audio per request or duration. Prices are on the models page.

#Way 1: MCP server

The server is https://bigmodel.io/mcp and authenticates with your API key (Authorization: Bearer <your key>). Set BIGMODEL_API_KEY in your shell first.

#Codex

Add to ~/.codex/config.toml:

[mcp_servers.bigmodel]
url = "https://bigmodel.io/mcp"
bearer_token_env_var = "BIGMODEL_API_KEY"
tool_timeout_sec = 300

Videos usually take 1–3 minutes; a longer tool_timeout_sec lets the tool wait for the result.

#Claude Code

claude mcp add --transport http bigmodel https://bigmodel.io/mcp \
  --header "Authorization: Bearer $BIGMODEL_API_KEY"

#Cursor

Add to ~/.cursor/mcp.json (or .cursor/mcp.json in the project):

{
  "mcpServers": {
    "bigmodel": {
      "url": "https://bigmodel.io/mcp",
      "headers": { "Authorization": "Bearer ${env:BIGMODEL_API_KEY}" }
    }
  }
}

Other tools that support Streamable HTTP MCP servers take the same URL and header.

#Tools

Links are valid for 24 hours. Have the agent download them into the project with curl -L -o <file> <link>, for example:

Use BigModel to make a 1024x1024 flat-illustration app icon and download it to assets/icon.png.

#Way 2: AGENTS.md note

Put this in AGENTS.md at the project root (CLAUDE.md for Claude Code); the agent runs the commands itself.

## Generating images, video and audio

Use BigModel.io. The key is in the environment variable `BIGMODEL_API_KEY`; never write it into code or commit it.
Result links are valid for 24 hours: download them into the project right away.

- Models: `curl -s "https://bigmodel.io/v1/models?type=all" -H "Authorization: Bearer $BIGMODEL_API_KEY"`; use models whose `architecture.modality` is image, video or audio.
- Image (returns when done):
  ```bash
  curl -s https://bigmodel.io/v1/images/generations \
    -H "Authorization: Bearer $BIGMODEL_API_KEY" -H "Content-Type: application/json" \
    -d '{"model":"<image model>","prompt":"<description>","size":"1024x1024","n":1}'
  ```
  Download `data[0].url`: `curl -L -o assets/<name>.png "<url>"`.
- Video (async task):
  ```bash
  curl -s https://bigmodel.io/v1/videos \
    -H "Authorization: Bearer $BIGMODEL_API_KEY" -H "Content-Type: application/json" \
    -d '{"model":"<video model>","prompt":"<description>","seconds":5}'
  ```
  Poll `GET https://bigmodel.io/v1/videos/<id>` every 10 seconds until `status` is `succeeded`, then download `output.url`. `failed` is not charged.
- Audio (async task): `POST https://bigmodel.io/v1/audio/generations`, used like video; poll `GET /v1/audio/generations/<id>`.
- Tell me the estimated cost before generating (prices: https://bigmodel.io/models); ask me first if it is over $1.

#Troubleshooting

Picked an image or video model in the model selector and got model_endpoint_mismatch. Switch the main model back to a text model and use the tools above for media.

The MCP tools return 401. The agent did not get BIGMODEL_API_KEY, or the header does not carry the key.

The video stays "still running". Videos can take a few minutes. Ask the agent to call get_video later, or raise wait_seconds and the tool timeout.

Insufficient balance. Video is billed per second and the balance has to cover the estimate when you submit. Top up.

#Related

• Image generation, Video generation, Audio generation
• Codex, Claude Code, Cursor
• Multimodal overview