Livepeer CLI

The local agent loop — for pipelines the MCP can't finish

Same network, same key, same models as the Livepeer Agent MCP — but the CLI owns the agent loop on your machine: internal polling for every job kind, no wall-clock ceiling, cost caps, session replay, and offline/alternate-LLM support. Reach for it when a job needs more than a handful of tool calls or more than a few minutes.

Install

npm install -g @livepeer/agent
livepeer setup   # paste your sk_… Daydream key

No npm access yet? The installer falls back to a hosted tarball automatically.

No key yet?

Grab one free at app.daydream.live, or try the keyless demo first via MCP (no CLI install needed).

Keyless demo →

First run

livepeer agent run "watercolor fox in studio ghibli style" --approval auto

Or open the interactive REPL: livepeer

Verb overview

Verb	What it's for
`livepeer agent run "<prompt>"`	One-shot or multi-step generation. The main entry point.
`livepeer`	Interactive REPL — conversational, session-scoped.
`livepeer config show --explain`	Resolved LLM provider + key status.
`livepeer config llm <provider>`	Switch the agent-loop LLM: `livepeer` (hosted, Daydream-billed), `gemini`, `claude`, `openai`, or `ollama` (fully offline).
`livepeer skills list / show / load / publish`	The agent's behavior layer — local files in `~/.livepeer/skills/`.
`livepeer session list / show <id> / replay <id>`	Every run writes a replayable JSONL ledger.
`livepeer cost [--json]`	Spend summary across all runs.
`livepeer forecast tool\|plan`	Estimate cost before you spend it.

When to prefer the CLI over the MCP

Same backend, same billing, same models — the difference is entirely in the agent layer: who runs the loop, what gets timed out, what persists across runs.

Multi-step pipelines with more than ~4 tool calls (image → animate → concat → mix → mux → export). The MCP's server-side submit_agent_task wrapper has a wall-clock ceiling; the CLI's local loop does not.
Long-running work — multi-shot films, multi-episode series, full ffmpeg finishing chains that take several minutes wall-clock.
Cost-bounded runs — --budget 25.00 enforces a hard runtime cap; nothing equivalent exists on the MCP path.
Scripted / batch / CI work — --json output, cron, GitHub Actions, looping across N brands or N products.
A different LLM — Ollama (fully offline), Gemini or Claude direct, custom OpenAI-compatible endpoints.
Session replay + cost audit after the fact — livepeer session / livepeer cost.
Live video-to-video streams — Scope scope_start/control/stop control is CLI-only today.

Prefer the MCP instead for a quick one-shot generation inside a chat you're already in, conversational refinement of a single asset, or composing with other MCP servers (Salesforce, Slack, etc. in the same prompt) — the CLI can't do cross-MCP composition. See the full write-up: MCP vs CLI, an honest comparison.

Typical recipes

6-scene illustrated story (character-locked)

livepeer agent run "Make a 6-scene illustrated story about an arctic explorer's
last journey. Studio Ghibli watercolor, soft pastel blues and amber. Use
gpt-image for the hero shot and kontext-edit with that hero as source_url
for the remaining 5 scenes." --approval auto --max-tool-calls 20 --budget 8.00

30-second cinematic clip with full finishing

livepeer agent run "30-second cinematic clip about a lone surfer at sunset.
6 keyframes via flux-schnell, animate each with seedance-mini-i2v 5s, voice
over with chatterbox-tts, score with lonely synth pads, concat with cuts,
mux, export youtube-landscape." --approval auto --max-tool-calls 30 --budget 3.00

Pricing before generating

Free, no charge for discovery: livepeer agent run "list the video models on the network and their pricing" --json, or call livepeer forecast plan "<brief>" to estimate a full pipeline before you run it.

← Get Started (MCP) · Skills · /start