Last updated: 2026-10-30
LangGraph is an MIT-licensed Python framework from LangChain for building stateful, multi-step AI agents as graphs, with human-in-the-loop approvals, persistent memory and token streaming. Its GitHub repository showed 44.2k stars and 7.6k forks on 2026-10-30, and the library is free to install and run.
About LangGraph
LangGraph is an open-source agent runtime and low-level orchestration framework from LangChain Inc. You describe an agent as a graph of steps that share state, then add loops, branches and approval points where the task needs them. The GitHub repository lists an MIT licence, 44.2k stars and 7.6k forks as of 2026-10-30, and the vendor states that the library is free to use.
The core idea is control. LangChain's own LangChain library is positioned for quick-start agents with any model provider, while LangGraph is the layer you reach for when the flow must be explicit. Its documented building blocks are human-in-the-loop interrupts that pause an agent until a person approves, built-in memory that keeps conversation history across sessions, and token-by-token streaming. The vendor says it supports single-agent, multi-agent and hierarchical control flows in one framework, and that it adds no overhead to your code. Treat that last point as a vendor claim, not a measurement.
It fits engineers who are already shipping agents and have hit the limits of a single prompt loop: support triage with a manager approval step, research agents that fan out and merge, or long-running jobs that must resume after a failure. Peers in AI orchestration take different trades. CrewAI leans on role-based crews, Mastra targets TypeScript teams, LlamaIndex starts from documents and retrieval, and n8n, Make and Zapier give a visual builder to teams that would rather not write code. If your question is which agent to buy rather than build, start with the best AI agents guide.
The library costs nothing, but running it in production is a separate decision. LangChain sells hosted deployment as LangSmith Deployment. The Developer plan is $0 per seat with one seat and does not include deployment. Plus is $39 per seat per month, adds unlimited seats and includes one free Serverless (Small) deployment, with extra deployments metered in LangChain Standard Units at $1.00 each. Enterprise pricing is custom and adds self-hosted and hybrid options. You can also run LangGraph on your own infrastructure at no licence cost, with models from providers such as Anthropic or OpenAI.
The honest downsides are the learning curve and the hosted bill. A graph with typed state is more code than a prompt and a loop, and the vendor itself calls LangGraph low-level. Tracing and evaluation live in LangSmith. Teams that want tracing elsewhere can look at the other observability tools in the category, such as ReasonBlocks (compared in ReasonBlocks vs Touchmark), Archal, Confident AI and Openlayer, and should read the three jobs of agent observability guide before committing. Durable workflows that are not LLM-specific may suit Inngest better, and OpenAI-only teams can compare the OpenAI Agents API. For the wider picture, see the wider AI tools guide.
Pricing
Library free under the MIT licence. Hosted deployment runs through LangSmith: Developer $0 per seat, Plus $39 per seat per month, Enterprise custom. Read from langchain.com/pricing on 2026-10-30.
| Tier | Monthly price | What it includes |
|---|---|---|
| LangGraph (open source, MIT) | Free | |
| LangSmith Plus (hosted deployment) | $39/mo | |
| LangSmith Enterprise | Custom |
Key Features
- Human-in-the-loop interrupts: Agents can pause at a defined step and wait for a person to approve or edit an action before the graph continues.
- Persistent memory: Built-in memory stores conversation history so an agent keeps context across separate sessions.
- Flexible control flows: The vendor documents single-agent, multi-agent and hierarchical designs built from the same low-level graph primitives.
- Native streaming: Token-by-token streaming shows agent reasoning and actions as they happen instead of after the run.
- LangSmith Deployment: Plus includes 1 free Serverless (Small) deployment for running stateful, long-running agents, with extra deployments metered per resource hour.
Pros
- MIT licence with no per-use fee, so the framework itself adds $0 to an agent project.
- Explicit approval points make it a fit for workflows where a wrong action is costly.
- Large open-source base with 7.6k forks on GitHub, so worked examples and community answers are easy to find.
Cons
- Low-level by design: a typed-state graph is more code than a prompt loop, so small projects pay a setup cost.
- The hosted route adds seat and metered costs, and the Developer plan lists no deployment at all.
- Tracing and evaluation sit in LangSmith, so full observability means a second product and a second bill.
Frequently Asked Questions
What does LangGraph cost, including the hosted option?
The LangGraph library is MIT-licensed and free. Hosted deployment is sold through LangSmith: Developer is $0 per seat and lists no deployment, Plus is $39 per seat per month with one free Serverless (Small) deployment, and Enterprise is custom. Anything beyond the plan is billed per LSU at $1.00.
Is LangGraph open source, and can a company use it commercially?
Yes. LangChain describes LangGraph as an MIT-licensed open-source library, and the GitHub repository shows the MIT licence. The MIT terms allow commercial use, so the main cost is engineering time and whatever model providers you connect.
Which tools should I weigh against LangGraph?
Take CrewAI if role-based crews match your task, LlamaIndex if document retrieval is the core job, and Mastra if your team writes TypeScript. n8n is the pick when non-developers need to edit the workflow visually.
How is LangGraph different from LangChain?
LangChain presents its own library as the quick way to start agents with any model provider. LangGraph is the lower-level runtime for explicit control flow, persistence and approvals. Start with LangChain for a prototype and move to LangGraph when the flow needs branches, pauses or resumable state.
How do I start building with LangGraph?
Read the Python overview in the LangChain docs and take the free Introduction to LangGraph course on LangChain Academy. The GitHub repository at langchain-ai/langgraph has examples. Plan a first graph with two or three nodes and one human approval step before adding more.
Top Alternatives
- LangChain: Pick LangGraph if you need explicit control flow and approvals; pick LangChain for quick-start agents with any model provider.
- CrewAI: Pick LangGraph if you want to design the graph yourself; choose CrewAI when role-based crews match how you think about the task.
- LlamaIndex: Pick LangGraph for agent control flow; choose LlamaIndex when document parsing and retrieval are the main job.
- Mastra: Pick LangGraph for a Python codebase; choose Mastra if your team works in TypeScript.
- n8n: Pick LangGraph if engineers own the agent code; choose n8n for a visual builder that non-developers can edit.