@pollo-ai/cli

Generate images and videos with Pollo AI from your terminal — or expose the same capabilities to an AI coding agent over MCP.

Install

npm i -g @pollo-ai/cli

Works on macOS, Linux and Windows. The package is a small Node shim; the actual binary ships in a per-platform package that npm picks for your machine, so nothing is downloaded by a postinstall hook.

Without Node (macOS and Linux):

curl -fsSL https://raw.githubusercontent.com/polloaiofficial/cli/main/install.sh | sh

Log in

pollo auth login

This opens your browser, you sign in to Pollo, and the CLI stores the resulting token in ~/.config/pollo/config.json with file mode 0600. The short-lived token refreshes itself before each request, so you won't be asked again until the long-lived one expires.

pollo account status

Shows who you are, your available credits, and your subscription. If this works, you're set up.

Generate

Pick a model, then read what it accepts before you build a request — parameters differ a lot between models:

pollo model list                        # every available model
pollo model list --type text2video      # only models with that capability
pollo model get google/nano-banana --fields

--type filters by capability: text2video, image2video, ref2video, text2image, image2image. Each model is listed once with its capabilities aggregated.

model get --fields returns each endpoint's fields ({name, type, description, values?, min?, max?}) plus cross-field constraints ({id, message, expr}) — already flat, no JSON Schema to parse.

Then submit and wait:

# Optional: estimate the credit cost first
pollo generate cost bytedance/seedance-2-0-fast --prompt "studio product video" --duration 5

pollo generate create bytedance/seedance-2-0-fast \
  --prompt "studio product video" --duration 5 --json     # -> { "taskId": "…" }

pollo generate wait <taskId> --json                        # -> result URLs

More examples:

# Text to image
pollo generate create openai/gpt-image-2 --prompt "a cozy cabin in a snowy forest" --json

# Image to image — a local path is uploaded for you, or pass an HTTPS URL
pollo generate create google/nano-banana --image ./photo.png --prompt "make it nighttime" --json

# Reference images (repeat the flag for several)
pollo generate create alibaba/happyhorse-1.1 --refs ./product.png --prompt "slow orbit around the bottle" --json

# A model with both video and image output needs --type
pollo generate create kling-ai/kling-v3-omni --type image --prompt "poster" --json

# Four outputs in one task, billed per output
pollo generate create bytedance/seedance-2-0-fast --prompt "studio product" --numOutputs 4 --json

pollo generate get <taskId> --json       # one-shot status check
pollo generate list --json               # your recent jobs

Notes on the flags:

• Every --name value you pass becomes an input field. Values are typed automatically: true/false → boolean, numeric strings → number, everything else → string. Repeat a flag to build an array.
• Image and video fields (--image, --images, --refs) take either a local file path or an HTTPS URL. Local files are uploaded as part of the generation; JPG and PNG only, so convert webp/heic first.
• --type is only needed when a model has both video and image output — the CLI resolves the endpoint on its own otherwise, and tells you when it can't.
• --numOutputs N (1–4) produces N outputs in one task. cost is billed per output, and get/wait list every one of them.
• --webhookUrl <url> gets you a callback instead of polling.
• --json prints the raw response. Without it, get/wait print a human summary of only the first output, so use --json when you asked for several.

When something fails

generate wait --json gives you a failMsg per output. Two cases worth distinguishing:

• Content moderation — change the prompt, not the model.
• Insufficient credits — the error carries a recharge link; retrying or switching models won't help.

Use as an MCP server

pollo mcp runs the CLI as a local stdio MCP server, so MCP-capable clients (Claude Code, Cursor, Claude Desktop, …) can drive Pollo directly. It reuses the login above — run pollo auth login once and the server picks up the same credentials.

Claude Code:

claude mcp add --scope user pollo -- pollo mcp

Cursor (~/.cursor/mcp.json or a project .cursor/mcp.json), Claude Desktop (claude_desktop_config.json), and other stdio clients take the same shape:

{
  "mcpServers": {
    "pollo": { "command": "pollo", "args": ["mcp"] }
  }
}

Tools exposed: pollo_account_status, pollo_list_models (pass format: "fields" to read a model's input fields), pollo_estimate_generation_cost, pollo_generate_image, pollo_generate_video, pollo_get_generation_status. The generate_* tools return {taskId, status} immediately — poll pollo_get_generation_status for the result URLs. Image inputs accept a local path or an HTTPS URL; there is no separate upload tool.

Teach an AI agent to use this

There's an agent skill that walks a coding agent through model choice, input preparation, cost estimation and polling:

npx skills add polloaiofficial/skills

Reference

pollo --help
pollo <command> --help

Global flags: --json (raw JSON), --no-color.

Credentials live in ~/.config/pollo/config.json (or $XDG_CONFIG_HOME/pollo/config.json) at mode 0600. pollo auth logout removes them.

License

MIT