# LiveGraph

> LiveGraph is a model-agnostic multi-agent orchestration platform built around a live,
> editable canvas: agents are nodes, routing is edges, and a run can be watched, interrupted,
> approved, and rerouted while it is executing — the engine re-reads the graph on every hop,
> so dragging an edge mid-run changes where the run goes next. Hosted at livegraph.ai.

## Product

- [Home](https://www.livegraph.ai/): what it is — watch agents think, steer them mid-run
- [Live demo](https://www.livegraph.ai/demo): run and reroute a real multi-agent graph in the browser, no signup
- [Pricing](https://www.livegraph.ai/pricing): free tier with an included model; bring-your-own-key or LiveGraph-hosted keys
- [Docs](https://www.livegraph.ai/docs): orchestration model, live vs pinned runs, approvals, automations
- [Changelog](https://www.livegraph.ai/changelog): shipped features and fixes, most recent first
- [Status](https://www.livegraph.ai/status): live service status and recent probes
- [Security](https://www.livegraph.ai/security): credential encryption, worktree isolation, allowlists, approval-gated writes
- [Automations catalog](https://www.livegraph.ai/templates): ready-made agent setups that instantiate into editable graphs
- [MCP connectors](https://www.livegraph.ai/mcp): connect agents to remote MCP servers, or publish a graph as an MCP server
- [Compare](https://www.livegraph.ai/compare): honest side-by-sides against the tools teams evaluate
- [Blog](https://www.livegraph.ai/blog): notes on live agent orchestration

## Comparisons

- [LiveGraph vs n8n](https://www.livegraph.ai/compare/n8n): n8n executes a persisted workflow snapshot and its canvas goes inert at run time; LiveGraph's canvas stays live and steerable while a run executes. An honest side-by-side.
- [LiveGraph vs Lindy](https://www.livegraph.ai/compare/lindy): Lindy hides the agent graph entirely — there's no canvas to open or steer. LiveGraph is the graph: visible, editable, and live while a run executes. An honest side-by-side.
- [LiveGraph vs Gumloop](https://www.livegraph.ai/compare/gumloop): Gumloop's flow canvas goes inert once a run starts — the documented answer to a stuck run is restarting it. LiveGraph's canvas stays live and steerable mid-run. An honest side-by-side.
- [LiveGraph vs Zapier Agents](https://www.livegraph.ai/compare/zapier-agents): Zapier Agents is the black-box end of the category — no orchestration graph to inspect or steer. LiveGraph is a live, editable agent canvas. An honest side-by-side.
- [LiveGraph vs LangGraph](https://www.livegraph.ai/compare/langgraph): LangGraph's checkpointing is the deepest intervention model in the field — but it forks state inside a topology frozen in code. LiveGraph rewires the topology itself, mid-run. An honest side-by-side.
- [LiveGraph vs Langflow](https://www.livegraph.ai/compare/langflow): Langflow's drag-and-drop canvas is where you build a flow — at run time it's a test surface, not a steering wheel. LiveGraph's canvas stays live and steerable mid-run. An honest side-by-side.
- [LiveGraph vs Dify](https://www.livegraph.ai/compare/dify): Every Dify run executes the published workflow version — the canvas you edit is a draft the run never sees. LiveGraph's engine re-reads the live graph on every hop. An honest side-by-side.
- [LiveGraph vs CrewAI](https://www.livegraph.ai/compare/crewai): CrewAI defines role-based agent teams in Python; its visual Studio is an enterprise tier that observes runs rather than steering them. LiveGraph puts the wiring on a live, editable canvas. An honest side-by-side.
- [LiveGraph vs Make](https://www.livegraph.ai/compare/make): Make runs activated scenarios on schedule — once an execution is in flight you can't even stop it from outside the editor. LiveGraph's canvas stays steerable mid-run. An honest side-by-side.

## Migrate

- [Migrate from Flowise](https://www.livegraph.ai/migrate/flowise): import Flowise agentflows onto the live canvas
- [Migrate from OpenAI Agent Builder](https://www.livegraph.ai/migrate/agent-builder): rebuild Agent Builder workflows (shutting down 2026-11-30)
- [Migrate from n8n](https://www.livegraph.ai/migrate/n8n): move n8n workflows onto a canvas that stays live during the run
- [Migrate from Langflow](https://www.livegraph.ai/migrate/langflow): move Langflow flows onto a canvas that stays live during the run
- [Migrate from Dify](https://www.livegraph.ai/migrate/dify): move Dify agents and workflows onto the live canvas
- [Migrate from CrewAI](https://www.livegraph.ai/migrate/crewai): move CrewAI crews onto a visible, steerable canvas
- [Migrate from Make](https://www.livegraph.ai/migrate/make): move Make scenarios onto the live canvas

## Embedding

A finished or in-flight run can be iframed into another page at
https://www.livegraph.ai/embed/runs/<run-id>?t=<share-token> — a chrome-less, read-only canvas showing
topology and per-hop status (never hop text). The token is the grant; no auth required.

## API

The REST API lives on the api host (https://api.livegraph.ai), not the web origin. Programmatic callers
authenticate with an `X-Api-Key` header carrying an `lgk_…` key, minted under
Settings → API keys (or POST /me/api-keys on a session). The key acts as its owner — every
graph role check applies unchanged, so a member's key can only do what that member can.

- Publish a graph as an MCP tool: a graph's owner assigns it a slug in graph settings,
  then POST https://api.livegraph.ai/mcp/serve/:slug is a single-POST-per-request JSON-RPC 2.0 endpoint
  (initialize, tools/list, tools/call, resources/read — no SSE). tools/list advertises
  exactly one tool, run_<graph-name>, taking {prompt, context?}; tools/call launches a
  pinned run, waits up to ~30s for a result, else returns a run://<id> resource URI the
  caller polls via resources/read.
- X-Api-Key (member) path: the caller needs at least viewer on the served graph; calls are
  rate-limited per caller+graph and bounded by the caller's monthly spend cap.
- Anonymous (public) path: only when the owner opts the slug into public access. No key
  needed; capped at 20 launches/day per client IP, 50 launches/day and 12/hour per graph,
  and $2/day of metered spend per graph — anonymous calls bill to the graph owner's
  credentials, and public tool output is marked as untrusted third-party content.

## For agents

- [Agent guide](https://www.livegraph.ai/agents.md): how an AI agent uses or integrates LiveGraph — MCP serving, the runs API, API keys
- [llms-full.txt](https://www.livegraph.ai/llms-full.txt): the expanded variant — every public page as a link list

## Data feeds

- [Model Radar](https://www.livegraph.ai/model-radar): weekly bench of newly-released LLMs on OpenRouter,
  measured on the same call path production agent graphs use — score, wall-clock latency,
  and estimated cost per eval call
- [Model Radar API](https://api.livegraph.ai/public/model-radar): the latest scan as JSON,
  CORS-open, no auth. GET /public/model-radar/scans lists historical scans; ?scan=<id> reads one

## Key differentiator

Generated, templated, or hand-built graphs all land on the same observable canvas — during a
run each hop re-resolves routing against the graph's current state, so an operator can drag an
edge and the next hop follows it. No opaque workflow surface.

## Optional

- [Privacy Policy](https://www.livegraph.ai/privacy): what is collected, how credentials are stored, how data flows to model providers
- [Terms of Service](https://www.livegraph.ai/terms): accounts, billing, agent autonomy, acceptable use