A CLI Packs Image, Video, and Audio into Agents—and the Comments Immediately Ask About the Bill

Chris Tate posted Vercel Labs' experimental project ai-cli on X. It doesn't train models, and it doesn't provide a chat interface. It puts text, image, video, and audio generation—plus a set of typed judgments—into terminal commands. Under the hood, it uses the Vercel AI SDK and AI Gateway to connect to hundreds of models. Installation is one line: npm i -g ai-cli. The repo requires Node.js 22+ and an AI Gateway key, or a key from the corresponding model provider. The repo is at github.com/vercel-labs/ai-cli.

The launch video lists entry points and model combinations such as Claude Code, Codex, and opencode, and demonstrates calling media generation from within an agent. The commands are straightforward too. ai image 'a cute dog', ai video 'a spinning triangle', ai text 'explain quantum computing', ai audio speak 'Thanks for trying ai-cli'. It can also transcribe audio, list models, and start an MCP server. For agents, the MCP path matters more than a human manually typing commands; an agent can call ai as a tool.

More interesting is the pipeline. ai image 'a dragon' | ai video 'animate this' can feed an image result directly into a video model. Uses like git diff | ai text 'explain these changes' are not new, but putting them in the same CLI makes images, video, audio, and text into commands that can be chained together. Someone commented that a pipeline sending an image into a video model is the cleanest part of the launch. That assessment is not exaggerated.

Models can be mixed and selected. -m accepts comma-separated model IDs, and ai image 'a sunset' -m 'openai/gpt-image-1,xai/grok-imagine-image,bfl/flux-2-pro' runs several at once. Environment variables can change the default model, and ai models lists available models along with price, context, and latency. The repo also provides ai video start and ai video status; for multi-minute tasks like video, you can submit first and poll later, so the terminal doesn't get stuck.

The questions in the comments are more specific than the launch copy. Some ask whether they need to paste a key for every model; the repo's answer is that either an AI Gateway key or a provider key works. Some ask whether they can see costs before a call, and whether retry loops will burn through the bill. The docs include timeout, exit code, concurrency, and retry parameters, but I didn't see a design for per-session spending limits. Some ask what to do when video generation takes too long; video start/status and --poll-timeout-ms are for exactly that scenario. Others ask whether the agent chooses a model per asset or whether each CLI has a fixed model. In practice, you control it with -m and AI_CLI_*_MODEL; defaults are set in environment variables. Whether the agent can choose depends on how the caller wraps it.

One obvious shift is that media generation goes from being a web tool to being another tool call for agents. It doesn't solve which model is better, and it doesn't decide for you which model to use. It solves routing and chaining: in the same terminal, entry points like Claude Code, Codex, and opencode can send off image, video, and audio tasks, with results landing in files or stdout. For developers, it removes a layer of glue code. For ordinary users, it removes copying and pasting between multiple products.

What really needs watching is default models, retry strategy, and bill control. The comments have already raised this. If an agent learns to retry video generation on its own, the bill will look much worse than in the text era. Vercel Labs has built the pipeline; next we need to see who installs the brakes first.