LangGraph

What is LangGraph?

LangGraph (LangChain) models an agent as a graph: nodes, edges, and explicit shared state, so you can write loops, branches, retries, and human-in-the-loop pauses instead of a single chat while-loop. The Python and JS libraries are free and open source.

You pay when you want LangChain to host or observe the thing. Managed hosting that used to be branded LangGraph Platform is now LangSmith Deployment.

As of September 2026, LangSmith Developer is $0 for one seat with 5,000 traces/month. Plus is $39/seat/month with 10,000 traces and one free small serverless deployment; extra traces, compute (LCU at $1.50), storage (LSU at $1.00), and deployment runs are usage-billed.

Enterprise adds hybrid and fully self-hosted control planes. You can also run an Agent Server in your own Docker/Kubernetes with your own Postgres and Redis.

LLM tokens are always a separate vendor bill (OpenAI, Anthropic, etc.).

A concrete walkthrough: human-approve a refund agent without losing the thread

Model three nodes: retrieve_order → draft_refund → human_gate → execute. The gate pauses the graph and stores state so a Slack approve can resume tomorrow — this is the reason to use LangGraph instead of a for-loop in a Lambda.

Develop that locally on the free library. Turn on LangSmith Developer traces while you debug the first 100 runs.

When support wants a URL and you do not want to babysit Docker, Plus ($39/seat) is the first plan that includes a small Cloud deploy. If refund payloads cannot go to LangChain's cloud, run the Agent Server in your cluster and keep Developer/Plus only for traces — or skip SaaS tracing.

Tokens for the retrieve/draft nodes still hit OpenAI/Anthropic either way.

Key features

• Graph control flow with durable state — cycles, retries, and human approval as first-class nodes
• Streaming and persistence for long-running agents, not just one-shot chains
• Works with LangChain models/tools but can be used without the rest of the kitchen sink
• LangSmith tracing/evals on a free Developer seat (5k traces) or Plus ($39/seat)
• LangSmith Deployment (Cloud) on Plus+: one small serverless deploy included, then usage
• Self-host the Agent Server if data cannot leave your VPC; Enterprise for a managed control plane in your cluster

How to get started

1. pip/npm install langgraph and write the graph locally — no credit card
2. Add LangSmith tracing when you need to see which node failed; Developer is enough to start
3. Deploy with `langgraph deploy` or the LangSmith UI only after Plus is justified (team + hosted runtime)
4. Self-host the server (Docker) if Plus's Cloud deploy is a data-residency problem
5. Budget model tokens separately; LangSmith never replaces the OpenAI/Anthropic invoice

LangGraph pricing

LangGraph library: free. LangSmith Developer: $0, 1 seat, 5k traces/month.

LangSmith Plus: $39/seat/month, 10k traces, 1 small serverless deploy included; extra usage on traces, LCUs ($1.50), LSUs ($1.00), and deployment runs. Enterprise: custom (hybrid / self-hosted platform).

LLM usage billed by your model vendor. Checked September 2026 on langchain.com/pricing.

Our take: Write the graph for $0. Buy Plus when you want LangChain to host it or when 5k traces are gone. Do not confuse $39/seat with model spend — a chatty graph on GPT will dwarf the LangSmith line.

Prices verified 2026-09. Plans change often — confirm on the official site above before you buy.

Use cases

• Production agents that must retry a tool call or wait for a human without losing state
• Multi-step research or coding loops you can audit node by node
• Teams already on LangChain who outgrew a linear chain
• Long sessions that need a checkpointer, not a stateless POST

Strengths and tradeoffs

LangGraph alternatives

• CrewAI: CrewAI is role-based crews (researcher, writer) with less explicit graph state — faster to sketch, harder to audit.
• AutoGen: AutoGen is conversation-of-agents; good for prototypes, looser than a LangGraph state machine.
• n8n: n8n is visual ops automation; use it to call an already-built agent, not to implement the agent's internals.
• Flowise: Flowise is drag-and-drop LLM chains if the team will not write Python/JS graphs.

Use LangGraph when control flow and durable state are the product. Use CrewAI to sketch role-based crews.

Use AutoGen for conversational multi-agent experiments. Use n8n when the work is wiring SaaS tools, not writing an agent runtime.

FAQ

Is LangGraph free to use?

LangGraph's plans: Open-source framework; LangSmith Plus $39/seat to host. Pricing and free-tier limits change over time, so check the official site above for the latest details.

What is LangGraph used for?

A graph runtime for agents — the library is free; hosted deploy is a LangSmith plan.

Who is LangGraph best for?

LangGraph is a good fit for production agents that must retry a tool call or wait for a human without losing state, or multi-step research or coding loops you can audit node by node.

What are the alternatives to LangGraph?

Commonly compared alternatives include CrewAI, AutoGen, and n8n, see the comparison above for how they differ.

How much does LangGraph cost?

The LangGraph framework is free. Hosted deploy and team tracing sit on LangSmith: Developer is free (5k traces), Plus is $39 per seat per month plus usage. Your OpenAI or Anthropic tokens are a separate bill.

What's the best LangGraph alternative?

CrewAI is the most commonly recommended swap: crewAI is role-based crews (researcher, writer) with less explicit graph state — faster to sketch, harder to audit. Which one is actually best depends on which of those tradeoffs matters more for what you're doing.

Is LangGraph free?

The framework is free and open source. You pay LangChain only for LangSmith (tracing/evals) and LangSmith Deployment (hosted agents). Developer is $0 with 5,000 traces. Plus is $39 per seat per month plus usage.

Do I need LangSmith to run LangGraph in production?

No. You can run the Agent Server on your own Docker/Kubernetes with your own database. LangSmith is how you get official hosted deploy, a UI for traces, and team features. LLM tokens are always billed by the model provider.