Why 8.5/10: Top pick for agent builders: MIT-licensed, actively maintained, and designed for the hard parts of agents (state, persistence, interrupts and human review). Points off for a steep learning curve and for the paid LangSmith platform being where hosting and observability live.
By ToolSift editorial. Last checked . Prices change often; "check site" means verify with the vendor.
What it does
LangGraph is an open-source library from LangChain for building agents as graphs: nodes are steps (a model call, a tool call, your own code) and edges decide what runs next. LangChain's overview describes it as "very low-level, and focused entirely on agent orchestration", and lists persistence (checkpointers and stores), fault tolerance, streaming, interrupts, time travel, memory and subgraphs as core capabilities. Interrupts are what make human-in-the-loop approval possible: an agent can pause, wait for a person, and resume from saved state.
The project is popular and busy. On 19 Oct 2026 the langchain-ai/langgraph GitHub repository showed about 43,500 stars, an MIT licence and a latest release of 1.2.16 (16 Oct 2026). LangChain's docs say companies including Klarna, Uber and J.P. Morgan use it. Documentation now lives on docs.langchain.com.
You can run LangGraph anywhere you run Python or JavaScript. LangChain's commercial platform, LangSmith, adds observability and evaluations, a Studio for debugging graphs, and managed deployment of agents. That is where the money is: the library is free, but production hosting and tracing through LangSmith are metered.
Against our runners-up: Dify gives you a visual, low-code builder with hosting, which is quicker for business teams; CrewAI offers a simpler role-based way to orchestrate several agents in Python. LangGraph sits lower down: it asks more of you up front but gives the most control over state, branching, retries and human approval, which is what production agents tend to need once the demo works. Many teams prototype elsewhere and move the hard parts to LangGraph.
Our verdict
LangGraph is the framework we would choose to build serious, custom AI agents in code. LangChain's docs call it a low-level orchestration framework and runtime for long-running, stateful agents, letting you mix hand-coded steps with LLM-driven steps in one graph. It is MIT-licensed, available in Python and JavaScript, and very actively maintained. The trade-off is that it is a developer tool: you write and own the logic, and the managed hosting, tracing and evaluation come from LangChain's paid LangSmith platform.
✅ Pick it if You're a developer who wants fine control over agent logic, persistence and human approval steps, with the option of managed hosting later.
⛔ Skip it if You want a no-code or drag-and-drop builder, or you don't want to write and maintain agent code.
Pricing: As of 19 Oct 2026: LangGraph library free and open source (MIT). LangSmith platform: Developer $0 (1 seat, up to 5k base traces a month, then pay as you go); Plus $39 per seat per month (up to 10k base traces a month, one free small serverless deployment, then usage metered in LangChain Standard Units at $1.00 per LSU); Enterprise custom, including self-hosted and hybrid options. Startups can apply for up to $10,000 in credits. Check site.
LangGraph is built and maintained by LangChain, Inc. in the open on GitHub (langchain-ai/langgraph).
What to watch
Learning curve: it is deliberately low-level. You design state, nodes and edges yourself, which is powerful but slower to start than drag-and-drop builders such as Dify.
Platform costs: the framework is free, but LangSmith deployment, tracing beyond the included volume, sandboxes and other services are metered in LSUs. Use the usage calculator on the pricing page before committing.
Fast-moving releases: frequent versions mean you should pin dependencies and read the changelog when upgrading.
Older links to langchain-ai.github.io/langgraph still work, but current docs are on docs.langchain.com, so check you are reading the latest version.
Model costs are separate: LangGraph orchestrates calls to model providers you choose and pay for directly.