Agent SDK

@livepeer/agent: CLI, library, and MCP server for building creative AI applications on Livepeer. AgentRunner, provider abstraction, ToolRegistry, and 40+ inference capabilities.

---

@livepeer/agent is the agent runtime package from the Storyboard project (livepeer/storyboard). It provides a provider-agnostic agent loop, 40+ Livepeer inference capabilities as callable tools, working and session memory, and an MCP server mode for integration with Claude Code, Cursor, and other MCP-compatible tools.

API key from daydream.live.

Installation

One-line installer (CLI on PATH + MCP auto-config):

curl -fsSL https://storyboard.daydream.monster/install.sh | bash
livepeer setup --non-interactive --api-key sk_live_xxx

The installer drops livepeer on your PATH and configures Claude Code / Cursor MCP entries automatically.

npm (once 1.0.0 is published):

npm install -g @livepeer/agent

Homebrew (macOS):

brew install livepeer/tap/livepeer-agent

CLI Verbs

The CLI exposes verbs for inference, file handling, finishing, and memory:

Inference:

# Single model call
livepeer run flux-schnell --prompt "a cat astronaut on Mars"
# → https://v3b.fal.media/files/.../image.jpg

# Animate a still image
livepeer cinemagraph ./portrait.jpg --motion "leaves drifting"

# JSON output (for scripts)
livepeer run flux-schnell --prompt "..." | jq '.url'

Finishing (ffmpeg-backed):

livepeer concat clip1.mp4 clip2.mp4 --transition crossfade-300
livepeer mux video.mp4 audio.mp3 --mode replace
livepeer caption video.mp4 --text "Hello" --start 0 --end 3
livepeer export video.mp4 --preset tiktok-portrait

Discovery:

livepeer models              # list all 40+ capabilities
livepeer models flux         # inspect a specific model
livepeer pricing flux-schnell
livepeer schema seedance-i2v

Memory:

livepeer memory pin "always use seedance-i2v-fast for animations"
livepeer memory recall "animation"
livepeer memory show

All verbs emit JSON when stdout is piped. Use --json to force JSON on a TTY or --no-json to force human output when piped.

Library API

import {
  AgentRunner,
  ToolRegistry,
  WorkingMemoryStore,
  SessionMemoryStore,
} from '@livepeer/agent';
import { GeminiProvider } from '@livepeer/agent/providers/gemini';

const provider = new GeminiProvider({ apiKey: process.env.GEMINI_API_KEY! });
const tools = new ToolRegistry();
const runner = new AgentRunner(
  provider,
  tools,
  new WorkingMemoryStore(),    // 800-token budget per turn
  new SessionMemoryStore(),    // queryable log across turns
);

const result = await runner.run({ user: 'generate an image of a mountain' });
console.log(result.finalText);

Importing individual verbs (without the full agent loop):

import { runRun } from '@livepeer/agent/cli/verbs/run';
import { runConcat } from '@livepeer/agent/cli/verbs/finishing';

const result = await runRun({
  name: 'flux-schnell',
  prompt: 'a mountain at dawn',
  apiKey: process.env.DAYDREAM_API_KEY,
});

Provider swap: change the LLM backend without changing application code:

import { ClaudeProvider } from '@livepeer/agent/providers/claude';
import { LivepeerProvider } from '@livepeer/agent/providers/livepeer';

// Claude directly
const provider = new ClaudeProvider({ apiKey: process.env.ANTHROPIC_API_KEY! });

// All LLM calls through Livepeer infrastructure (one API key)
const provider = new LivepeerProvider({ apiKey: process.env.DAYDREAM_API_KEY! });

MCP Server Mode

@livepeer/agent runs as an MCP server over stdio, exposing all 40+ inference capabilities as MCP tools to Claude Code, Cursor, and any other MCP host:

livepeer mcp

Add to Claude Code (~/.claude/claude_desktop_config.json):

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

The one-line installer configures this automatically. Once connected, Claude Code can call livepeer tools directly: run flux-schnell, cinemagraph, concat, detect, and all other CLI verbs are available as MCP tool calls.

Configuration

Credentials live in ~/.livepeer/credentials:

DAYDREAM_API_KEY=sk_live_xxx
LIVEPEER_SDK_URL=https://sdk.daydream.monster

Environment variable overrides take highest precedence: LIVEPEER_API_KEY or DAYDREAM_API_KEY, and LIVEPEER_SDK_URL.

livepeer setup show    # print current config (key redacted)
livepeer version
livepeer update

Skills

Skills are markdown system prompt files that shape agent behaviour for specific tasks. The public registry is browsable and installable from the CLI:

livepeer skills list
livepeer skills install cinemagraph
livepeer skills install --all
livepeer docs "animate portrait"    # search registry by keyword

Installed skills live in ~/.livepeer/skills/*.md and are loaded automatically by the agent runner when relevant.

@livepeer/agent is under active development through multiple phases (C8 is the current stable milestone). The CLI verb argument shapes and output shapes are frozen at C8; later phases (D–L) add new verbs without breaking existing ones.

Related Pages

Storyboard

The reference application built on @livepeer/agent and @livepeer/creative-kit.

MCP and Livepeer

Connecting Claude Code, Cursor, and other MCP tools to Livepeer.

Agents Overview

Eliza integration and BYOC agent patterns.

AI Pipelines

The underlying inference endpoints the agent SDK routes through.