For Engineering Leaders

Ship media-creation features without owning the stack

One Daydream key. One agent runtime. One envelope for AI generation, deterministic finishing, and external publishing — across 60+ models, self-hosted Docker tools, and bridged SaaS. Three consumer surfaces (CLI, MCP, embedded SDK) on the same network underneath.

4-minute read·Livepeer Network·Get started with MCP →·Full docs →

See it in one click

Proven playbooks, remixed into your own chat. Pick one, edit the brief, hit send. Browse all templates →

heroTry without a key

Image or text → real 3D-printable CAD model (STL)

$0.05–$0.50 per part (priced per render + per refine pass)

Remix ▸
heroTry without a key

Format Bridge — convert any asset to the format your tool needs

$0.01 per conversion (deterministic format-convert cap)

Remix ▸
hero

Supervision — per-frame, per-instance video CV (track / blur / zone / heatmap / annotate)

$0.05–$3.00 per clip (priced at $0.01/s of source video × cap count)

Remix ▸

What the Agent SDK is

@livepeer/agent is a TypeScript runtime that wraps the Livepeer decentralized orchestrator network in a standard agent loop (LLM → tools → memory → outcome). It bundles five building blocks your team would otherwise write themselves:

  • LLM providers — Livepeer-hosted Gemini (default, billed via the user's Daydream key), Claude, OpenAI, Gemini direct, Ollama, custom OpenAI-compatible.
  • Tool registry — 21 built-in tools covering image / video / audio / 3D generation, canvas operations, memory, web access, file I/O.
  • Skills — markdown files that shape agent behavior at runtime (style guides, brand voices, workflow patterns). Ranked + injected into the system prompt per turn.
  • Memory + session ledger — JSONL per run for cost tracking, replay, preview generation, and outcome scoring.
  • Approval gate — interactive / auto / dry-run consent before paid tool dispatches.

All five run on top of one network call:POST /inference {capability, input}. That envelope dispatches to AI inference (fal.ai / Gemini), deterministic Docker tools (ffmpeg / OpenCV / YOLO), or MCP-bridged external services (Slack / Drive / Linear / GitHub) — by capability name. No client-side branching on "type of work."

The full media-creation lifecycle, in one envelope

Most teams stitch together five separate vendor accounts to ship a creative feature: an LLM provider, an image generator, a video model, a TTS service, an asset CDN. Then they build the glue: ffmpeg fleet, asset storage, payment splits, observability across five status pages. Livepeer collapses all of it.

           PLAN              GENERATE             FINISH               PUBLISH
           ────              ────────             ──────               ───────
LLM:      gemini-text   →   image: flux/gpt-image   →   ffmpeg-concat   →   slack-send-file
                            video: seedance/veo          ffmpeg-mux           drive-upload
                            audio: chatterbox/music      opencv-smart-crop    publora-schedule
                            3D:    tripo                 yolo-segment         figma-export

                                ALL THROUGH ONE BEARER, ONE ENVELOPE
                                ── billed per-call to the user's Daydream key ──

Generation is half the story. The finishing pass (concat, mux, smart-crop, caption-burn, platform export) and the publishing pass (Slack / Drive / multi-platform scheduling) live in the same envelope. Your agent never leaves the network mid-pipeline. The result your users see is a finished deliverable, not a list of fal.media URLs.

Built to flex, personalize, and learn

Flexible

Three consumer surfaces — CLI, MCP server, embedded SDK — share one backend. Swap the LLM with a config flag (livepeer config llm gemini). Swap an AI capability's upstream provider via config-only edit toCAPABILITIES_JSON— no SDK release. Open-source orchestrators mean no vendor lock-in for compute.

Personalizable

Write a markdown skill and your agent obeys it next turn. Train a LoRA on your brand assets — every generation references it via lora_id. Fork any public agent manifest from the marketplace and re-sign under your own id. Skills + LoRAs + agents cross-sync across devices via the per-user mirror.

Self-learning

Every turn is scored — was a tool called, was the user happy, did the pipeline finish. The skill ranker bumps skills that produced good outcomes; the model router learns actual per-model latency and weights faster models on the next call. Skill autogenesis can author candidate skills from successful sessions, surfaced for one-click promotion.

The compounding shape: your installation gets better the more your team uses it. Skill rankings, model timings, brand LoRAs, agent reputations — all per-user, all persisted, all carrying forward. A campaign run on day 1 takes 4 hours; a similar campaign on day 60 takes 8 minutes because everything is pre-tuned.

Get started in under 10 minutes

Four supported paths. Pick by where your engineering team lives.

The bottom line for your stack decision

Vendor accounts to provision1 Daydream key
Time to wire a new AI modelMinutes — config-only registration
Finishing pipelineBuilt into the same envelope
Lock-inDecentralized orchs compete for the work
Real-time video transformationFirst-class — LV2V on Scope orchs
Embeddable runtime@livepeer/agent — your UI, our network
Time to first production feature2-4 weeks (CLI) / 4-8 weeks (embedded)