LangGraph

LangGraph: the build option, with a managed runtime if you want one

LangGraph is the clearest "build" choice on this site. LangChain's page title sells it on reliability¹; the overview doc is more technical, describing orchestration primitives and a runtime aimed at stateful agents with long lifetimes². For an engineering leader, the appeal is control: the framework does not prescribe prompts or agent architecture³, and the team decides how every step behaves. Around the framework, LangChain sells LangSmith, which adds tracing, evaluation and a managed deployment service. LangGraph ranks third here.

This profile looks at what a team takes on when it builds on LangGraph, what LangSmith takes back, and what is left for the buyer to handle.

The framework and its license

The license file is the MIT License⁴, and the FAQ describes LangGraph as an MIT-licensed open-source library that is free to use⁵. The repository showed 44.7k stars when checked⁶. Its short description reads "Build resilient agents"⁷, and it describes the project as a low-level orchestration framework for building stateful agents⁸.

For a platform team, the license question is settled quickly: the MIT grant is the whole of it^4,9. Dify's modified license, by comparison, restricts multi-tenant operation without written permission¹⁰.

LangGraph stands apart from its sister library. The docs say it does not require LangChain¹¹, and the repository notes that LangChain Inc builds it while it can be used without LangChain¹². The FAQ adds that LangGraph will not add overhead to the user's code¹³.

How agents are modeled

Workflows are defined as graphs of shared state, nodes and edges¹⁴. Hand-coded nodes and model-driven nodes can sit in the same graph¹⁵, which lets a team keep deterministic business logic in code while a model handles the open-ended steps. This is the design choice that separates a framework from a no-code builder: the engineer draws the line between code and model, node by node.

Persistence is central. The framework checkpoints state and resumes after failures or pauses¹⁶. Checkpointers save a thread's graph state, giving short-term memory inside one thread¹⁷, while stores persist application data outside the graph state and across threads¹⁸. Memory covers short-term thread state and longer-term memory across sessions¹⁹.

Human oversight is built into the model. A run can pause, let a person inspect or edit state, and then continue²⁰, and the home page describes human-in-the-loop checks that steer and approve agent actions²¹. Streaming covers tokens, state updates and other events as the graph runs²², with native token-by-token streaming listed on the home page²³.

Developer tooling

Installation is a single package command²⁴, and the docs point to a full calculator agent in the quickstart²⁵. The LangGraph CLI provides a local development server²⁶. LangSmith Studio is a free visual interface for developing and testing agents²⁷; it connects to an agent running locally and shows each step it takes²⁸. With tracing disabled in the local environment file, no data leaves the local server²⁹.

That last detail is useful to security reviewers during early prototyping. A team can explore agent behavior on a laptop without sending traces to a hosted service.

LangSmith: where the "buy" part comes in

LangGraph Platform was renamed LangSmith Deployment in October 2025³⁰. LangChain describes it as the managed service for running agents at scale in production³¹, and it accepts agents built with any framework³². The deployment docs name Claude Agent SDK, Strands, CrewAI and AutoGen among the supported frameworks³³, and a Google ADK agent can be deployed as a LangGraph³⁴.

LangSmith covers tracing, evaluation, prompts and deployment across frameworks³⁵. LangSmith Engine reads LangGraph agent traces, detects issues and proposes fixes³⁶. The agent registry supports A2A, MCP and Agent Protocol natively³⁷ and manages all agents centrally with versioning³⁸. Assistants can be shared, modified and reused³⁹. Webhooks trigger external systems⁴⁰, crons run agents on a schedule⁴¹, and an agent can be exposed as an MCP server⁴².

Deploying to LangSmith Cloud starts from a GitHub repository⁴³, public or private⁴⁴. A new deployment is created from a button in the console⁴⁵ and may take about 15 minutes⁴⁶. The Studio button opens the deployed graph for testing⁴⁷, the API URL is copied from the deployment details⁴⁸, and the deployment is called from Python through the LangGraph SDK⁴⁹.

Integrations

LangGraph works on its own⁵⁰. For model and tool integrations, LangChain points to its LangChain library, described as the agent framework with integrations for models, tools and agent loops⁵¹. The home page pitches building and shipping fast with any model provider⁵².

LangChain does not publish a single business-system connector count for LangGraph on the pages opened. That fits the framework model: connectors to a CRM or an ERP are code the team writes or pulls from the LangChain ecosystem. Nexus, by contrast, lists more than 4,000 integrations and connects them during the first build^53,54.

Security and identity

LangChain's footer offers a Trust Center link⁵⁵; when we saved that page, only its heading came through⁵⁶. SOC 2, ISO 27001, HIPAA and GDPR are not named on the LangChain pages opened for this review. A buyer should ask LangChain for its reports directly; their absence here reflects what loaded, and may not describe what LangChain holds.

On identity, the Enterprise plan includes custom SSO, ABAC and RBAC⁵⁷. Below Enterprise, the pricing table lists Google and GitHub as SSO options⁵⁸. Enterprise procurement includes an infosec review⁵⁹.

Authentication for deployed agents is the customer's job by design. LangSmith provides an auth system that can integrate with most schemes⁶⁰. Authentication is handled by the customer's own authenticate handler and authorization by its own handlers⁶¹. Custom auth is supported on all plans⁶², and there is no default authentication⁶³. The handler can also accept LangSmith API keys alongside custom auth⁶⁴. Authorization checks privileges and roles per resource⁶⁵.

For a platform team this is flexibility and responsibility at once. It can wire LangGraph agents into the company identity provider exactly as it likes, and it must do that work before an agent faces real users.

Governance

Security controls on the pricing page list organization roles of User and Admin⁶⁶ and role-based access control⁶⁷. Enterprise includes custom seats and workspaces⁶⁸. A seat is a distinct user in the organization, invited users included⁶⁹.

Agent authorization is listed as beta⁷⁰. Governance beyond these controls, such as approval rules for when an agent may act or a release gate before production, is something the team builds into its graphs with the human-in-the-loop features²⁰.

Hosting models and data

LangSmith Deployment offers Cloud, BYOC or the customer's own infrastructure⁷¹. In Cloud, LangChain hosts the control plane, data plane and databases⁷². In the hybrid option LangChain hosts the control plane and the customer hosts the Agent Servers and their data plane⁷³. Self-hosted means the customer runs both the control plane and the Agent Servers⁷⁴, and that is the path for full data residency or air-gapped setups⁷⁵. The pricing page puts self-hosted data in the customer's VPC⁷⁶ and cloud data in LangChain's cloud in the US or EU⁷⁷.

Running LangSmith on your own infrastructure requires the Enterprise plan plus a self-hosting add-on⁷⁸. That setup keeps trace and operational data in the customer's environment⁷⁹ and can run in the customer's own Kubernetes cluster⁸⁰. A trial license key for self-hosting requires contacting sales⁸¹.

Two details matter for the hybrid model. Agent Servers send traces according to an environment variable⁸², and leaving it unset sends traces to the default GCP US endpoint⁸³. A team with a residency rule should set it deliberately.

On data use, LangChain says LangSmith does not train models on customer data⁸⁴, and traces, prompts and outputs stay private to the customer's organization⁸⁵. Base traces are kept for 14 days⁸⁶ and extended traces for 180 days⁸⁷.

Pricing

Pricing is public and granular. The Developer plan is free per seat⁸⁸ with up to 5,000 base traces a month before pay-as-you-go⁸⁹. Plus costs 39 dollars per seat per month⁹⁰ with up to 10,000 base traces⁹¹, and lets a team add unlimited seats at that price⁹². Plus includes one free small serverless deployment⁹³. Enterprise is custom priced⁹⁴ and invoiced annually upfront⁹⁵; Developer and Plus traces are billed monthly in arrears⁹⁶.

Usage is metered in LangChain Standard Units at 1 dollar each⁹⁷. Runtime compute on deployments is 0.0675 units per vCPU hour⁹⁸, and cloud deployments are billed on the compute and database resources each one requests⁹⁹. Tuned evaluators are billed only when they run successfully¹⁰⁰, with one evaluator type metered at 0.015 units per run¹⁰¹. Sandboxes include 8 units a month¹⁰². Startups can receive up to 10,000 dollars in credits¹⁰³.

For budgeting a build, the license lines are small. The engineering time to design graphs, write connectors, set up auth and run the stack is the larger number, and it does not appear on LangChain's price list. Nexus prices per role automated, with consultancy support included on a software pricing model¹⁰⁴, which is the main reason it sits above LangGraph for teams short on engineers.

Stated limits

The pricing page marks tuned evaluators as a public beta for Plus and Cloud Enterprise in the US¹⁰⁵, and agent authorization as beta⁷⁰. Sandboxes are limited to 10¹⁰⁶. A support SLA is listed on the Enterprise plan¹⁰⁷. Persistent checkpoints support payloads up to 25 MB¹⁰⁸. In the persistence docs, the in-memory saver does not persist between restarts¹⁰⁹, and unbounded checkpoint growth is listed as a known issue¹¹⁰. Each limit is written down in the docs or on the pricing page, which helps an engineering team plan around it.

Who uses it

LangChain says developers trust LangGraph to build reliable agents¹¹¹ and that companies shaping the future of agents use it¹¹². The repository names Klarna, Replit and Elastic among its users¹¹³, and the home page carries a quote from Ally about multi-actor agentic workflows in its tech labs¹¹⁴. The plans are aimed at solo users getting started¹¹⁵, teams building and deploying agents¹¹⁶, and teams with advanced hosting, security and support needs¹¹⁷. The home page positions it for building agents with low-level control¹¹⁸.

Build, buy or both with LangGraph

To build, a team needs only the free library⁵, the local development server²⁶ and Studio, which keeps all data on the local server when tracing is off²⁹. To buy, it moves to LangSmith Cloud, where LangChain hosts the control plane, data plane and databases⁷²; Plus includes one free small serverless deployment⁹³, and Enterprise adds an infosec review⁵⁹ and a support SLA¹⁰⁷. To do both, it runs the hybrid model, with LangChain's control plane and the customer's own Agent Servers⁷³.

LangSmith Deployment also accepts agents from other frameworks³², which makes it a possible runtime for a company whose teams build in more than one. Billing on Cloud follows the compute and database resources each deployment requests⁹⁹, so cost scales with the infrastructure the team chooses.

Verdict

LangGraph is the best fit on this site for an engineering team that has decided to build and wants the framework to stay out of its way. It is MIT licensed⁴, model-agnostic⁵², and backed by a managed runtime that also supports hybrid, self-hosted and air-gapped deployment^71,75. It ranks third because the build stays with the customer: connectors, authentication⁶³ and governance policy are the team's own work, and no certifications were named on the pages that loaded. Teams that want an agent in production on a known date, with that work done for them, will find Nexus a closer match.