LangGraph vs CrewAI vs OpenAI Agents SDK: Three MIT Libraries, Three Separate Bills

Updated November 10, 2026. Versions, licenses, prices, and model defaults below were read that day from PyPI, npm, GitHub, and the vendor docs.

LangGraph, CrewAI, and the OpenAI Agents SDK are three harnesses you import. Each one is MIT. None of them charges for the import. The money starts when you want someone else’s traces, a hosted runner, or the model on the other side of the call.

The short version: use LangGraph when the run has to resume. Use CrewAI when the work is a crew of roles. Use the OpenAI Agents SDK when you want the smallest loop and you will host it. The rest of this page is the versions, the bills, and the places those three sentences break.

Which harness should you ship: LangGraph, CrewAI, or the OpenAI Agents SDK?

Ship the harness whose failure mode you can already name.

LangGraph is a graph of state, nodes, and edges. The overview calls it a low-level orchestration runtime for long-running, stateful agents, and it says the library does not prescribe prompts or an agent architecture. A checkpoint is how a paused or crashed run continues. That is the reason to take on the graph.

CrewAI is a crew of role-playing agents, plus Flows for event-driven steps. Version 1.15.39 runs on Python 3.10 through 3.13. The open-source library is the code. CrewAI AMP is the hosted studio, and its free plan stops at 50 workflow executions a month.

The OpenAI Agents SDK is an agent, a runner, tools, handoffs, guardrails, sessions, and a trace. The SDK guide says your server owns deployment, tools, state, and approvals. The same page separates that from the Agents API, which is a managed harness inside OpenAI’s service.

A framework comparison that only sorts release tags will not make this choice. Nineteen of those tags, read November 9, are on Best AI agent frameworks. What people actually downloaded through November 8 is Which AI agent framework do people install?. The workshop around the model, in the abstract, is What is an AI agent harness?.

What are the versions, licenses, and languages on November 10, 2026?

The version numbers do not line up, even inside LangGraph. GitHub’s latest-release endpoint, PyPI, and npm were three different artifacts on the morning of November 10.

The star and version rows are a reading of public APIs on November 10, 2026. The LangGraph releases response carried HTTP Date Tue, 10 Nov 2026 11:45:33 GMT. The PyPI reads carried 11:45:44 GMT. The numbers are licensed CC BY 4.0. Credit AI Agents Info if you reuse them.

Treat “latest” as a question about which artifact. A blog that says LangGraph is “on 0.4.14” is quoting the SDK tag. A blog that says it is “on 1.7.3” is quoting npm. The package pip install langgraph resolved to 1.2.20.

What does each library cost once you leave the README?

The license fee is zero. Three other bills are printed.

A buyer who hears “we’ll use LangGraph” still has to ask which of three things that means: the MIT library with your own Postgres, a LangSmith Developer seat, or Plus at $39 per seat so Deployment is on the table. A buyer who hears “CrewAI” has the same split between pip install crewai and an AMP workspace that includes 50 executions. A buyer who hears “OpenAI agents” has to ask whether the process is yours or the Agents API’s.

Model tokens sit on top of all three. This page does not reprint token prices. The harness that spends them, for the two coding products, is compared in Claude Code vs Codex pricing.

Which models can the harness call?

LangGraph does not pick a model. The overview says docs examples often use LangChain for models and tools, and that LangGraph does not require LangChain. You call a model from a node. There is no default model string on that page.

CrewAI’s LLM page, read the same day, says it has native SDK integrations for OpenAI, Anthropic, Google’s Gemini API, Azure, AWS Bedrock, and Snowflake Cortex. Every other provider goes through LiteLLM. The examples printed on that page include openai/gpt-5.6-terra, anthropic/claude-sonnet-4-6, and gemini/gemini-3.6-flash. Those are the page’s examples. The same page tells you to check the provider’s catalog because models are added and retired. The testing page is narrower: crewai test defaults to 3 iterations and gpt-6-luna, and it says the only provider for that command is OpenAI.

The OpenAI Agents SDK is explicit. If the agent does not set a model, the models page says the SDK uses gpt-5.6-luna with reasoning.effort="none" and verbosity="low". Set model="gpt-5.6-sol" when you want the frontier settings that page describes. Non-OpenAI providers start with the SDK’s own integration points. LiteLLM and Any-LLM are in the SDK as best-effort beta adapters. A handoff can carry a different model than the agent that passed it the turn. Hosted multi-agent on the Responses API is marked experimental on that page, and it is a different mechanism from a handoff.

How do tools and MCP attach?

All three can call tools. MCP is not one switch with the same owner.

LangGraph’s unit of work is a node. You pass tools into the node that should call them. The MCP page in the LangChain docs is MCPAdapter, and it requires langchain[mcp]>=1.4.0. The page says that namespace is in beta. Its quickstart builds a LangChain agent with create_agent, then hands that agent the tools list_tools() returned. Transports on that page are an in-process server, a script over stdio, and an http or https URL over streamable HTTP. Separately, LangSmith’s Agent Server exposes a deployed graph as MCP at /mcp, over streamable HTTP, for protocol revisions 2024-11-05 through 2026-07-28. The server page says that needs langgraph-api>=0.2.3 and langgraph-sdk>=0.1.61. Calling an MCP server from a graph, and publishing the graph as an MCP server, are two features.

CrewAI puts MCP on the agent. The MCP overview documents a mcps field, string references for a remote URL or a named catalog server, and structured configs. Transports listed there are stdio, server-sent events, and streamable HTTPS. The simple path needs the mcp library. The adapter path is crewai-tools[mcp]. AMP’s custom-MCP guide is streamable HTTP only, with no auth, a bearer token, or OAuth 2.0. The pricing page says Basic and Enterprise can both export a workflow as an MCP server.

The OpenAI Agents SDK documents four MCP shapes on its MCP page: HostedMCPTool, so OpenAI’s Responses API calls a public server; MCPServerStreamableHttp; MCPServerSse; and MCPServerStdio. Tracing records the list-tools call and the tool calls. Hosted MCP and a server your process opens are different trust boundaries. The integrations page says to use hosted MCP for a public server that fits the platform’s trust model, and a local or private server when your runtime should own the connection.

Tool use is the general name for the call. The JSON you send on every turn has its own cost, measured on Why tool JSON costs more than your system prompt.

Where do traces and evals live?

A trace tells you what the run did. An eval tells you whether a change helped. These libraries do not put both in the same box.

LangSmith’s pricing page includes tracing, monitoring, and online and offline evals on the observability product. Tuned Evaluators are priced at 0.015 LSU per successful evaluation run. The page says they are in public beta for Plus and for Cloud Enterprise in the US, and that skipped or failed evaluations are not billed. LangSmith Engine, on Plus, reads traces and proposes fixes. That is a LangChain product on top of the library, not a function inside langgraph.

CrewAI’s testing page is a CLI: crewai test, defaulting to three iterations and gpt-6-luna, OpenAI only. It prints task scores. AMP’s pricing comparison lists tracing, OpenTelemetry, LLM testing, guardrails, and human-in-the-loop input on both Basic and Enterprise. The open-source test command and the AMP checkboxes are different products.

The OpenAI Agents SDK turns tracing on by default. The tracing page says a run records generations, tool calls, handoffs, guardrails, and custom spans, and that you can disable it with OPENAI_AGENTS_DISABLE_TRACING=1 or a run flag. Sensitive inputs and outputs are included unless you turn trace_include_sensitive_data off. Grading those traces is the platform page Evaluate agent workflows: start with trace grading, then move to datasets and eval runs. That grader is not a function this page found inside the SDK.

How to build the tasks, the isolation, and the grader yourself is How to build an eval harness for AI agents. How to read a harness score without mixing boards is How do you evaluate an AI agent harness?.

Where does the run resume after a crash?

This is the sharpest split, and it is why “they all have memory” is a useless sentence.

LangGraph’s persistence page gives you two stores. A checkpointer saves snapshots of one thread’s graph state, for conversation continuity, human-in-the-loop, time travel, and fault tolerance. A store saves key-value data across threads. InMemorySaver and MemorySaver sit in RAM. The page says a process restart drops them. Production is PostgresSaver or AsyncPostgresSaver, and the thread_id has to stay under 255 characters. SqliteSaver is the local-file option. Agent Server, if you deploy there, handles the checkpointer for you and still requires PostgreSQL for threads, runs, assistants, crons, and the memory store. The checkpointer itself can be Postgres, MongoDB, or custom.

CrewAI Flows keep state on the flow object while it runs. @persist writes that state so a later kickoff can resume or fork. The default named on the page is SQLiteFlowPersistence. kickoff(inputs={"id": ...}) resumes under the same id. kickoff(restore_from_state_id=...) forks a new id from the latest snapshot. That is flow state. It is not the per-node graph checkpoint LangGraph describes.

The OpenAI Agents SDK session stores conversation items and prepends them on the next turn. The Python page’s built-in list is SQLiteSession, AsyncSQLiteSession, RedisSession, SQLAlchemySession, MongoDBSession, DaprSession, OpenAIConversationsSession, plus wrappers for compaction and encryption. If a run pauses for approval, you resume it with the same session. A session cannot be combined, in the same run, with conversation_id, previous_response_id, or auto_previous_response_id. Those are the server-managed continuation options. A session replays history into the next call. It does not restore a half-finished graph node.

If a human has to approve a tool, LangGraph pauses the graph, CrewAI AMP lists human-in-the-loop on the pricing page, and the OpenAI SDK pauses on an interruption you approve and then resume. The glossary entries are human in the loop and approval gate.

Where does production hosting sit?

LangGraph runs in your process. You bring the checkpointer. LangSmith Deployment is the managed runner, and the pricing FAQ says Plus includes one free small serverless deployment, meant to scale to zero, with dedicated deployments for a customer-facing agent. Enterprise can be fully managed, hybrid, or self-hosted. Developer, at $0, is the tracing seat. The “which plan” answer on that page puts Deployment on Plus.

CrewAI runs in your process when you use the library. AMP is the managed path: Basic includes infrastructure on the comparison table, and Enterprise is the line that says cloud, your VPC, or your own infrastructure. The pricing page does not print a region list or a per-execution overage.

The OpenAI Agents SDK runs in your process. The SDK guide’s words are that your server owns deployment, tool implementations, state storage, and approval decisions. Sandbox agents are available when the agent needs a container, files, and commands. That sandbox is still an SDK feature you configure. The Agents API is the product where OpenAI runs the harness.

Coding agents that already own a shell are a different shelf. Claude Code, Codex, and the Claude Agent SDK are on it. So are Google ADK and Pydantic AI, if the constraint is Google Cloud, Go, or typed Python outputs. Those five are outside this comparison on purpose. The harness table is the wider list. The harnesses hub is the shelf.

Which should a builder pick, and which should a buyer pay for?

Builder, the run must survive a restart or a person who answers tomorrow. LangGraph, with PostgresSaver or Agent Server. Budget a LangSmith Plus seat only if you want their deployment and their evals. A Postgres you already operate does not require that seat.

Builder, the spec is a team of specialists and the language is Python 3.13 or older. CrewAI. Stay on the library until a non-engineer needs the studio. AMP Basic’s 50 executions are a ceiling you can hit in a week of demos.

Builder, the product is a Python or TypeScript service you already deploy, and the control flow is handoffs plus a few tools. The OpenAI Agents SDK. Set the model on purpose if gpt-5.6-luna is the wrong cost. Turn sensitive trace data off before the first production run. If the organization is on Zero Data Retention, tracing will not be there.

Buyer, a vendor’s proposal says “open source, so the platform is free.” Ask for the seat, the execution cap, and the model invoice as three lines. LangSmith Plus is $39 per seat before traces past 10,000. CrewAI Enterprise has no public dollar figure. The OpenAI SDK’s platform fee is the model bill, plus a separate decision if they meant the Agents API.

Buyer, the proposal leads with GitHub stars. On this pass CrewAI had 60,723 and LangGraph had 44,746. From October 10 through November 8, 2026, the download page on this site counted 55,093,579 PyPI downloads for langgraph and 2,629,769 for crewai. Stars and installs answered different questions. Neither number is a quality score.

What do the star counts and the npm name get wrong?

Three mistakes show up in other comparison posts. Each one was false on the sources read today.

The first is one version number for LangGraph. Core Python, the Python SDK tag, and the npm package were 1.2.20, 0.4.14, and 1.7.3. Shipping a lockfile against “LangGraph latest” without saying which registry will not reproduce.

The second is npm install crewai. That name resolved to 1.0.1 from August 21, 2024, with a repository field of git+https://github.com/jaafarskafi1/crew-js.git. A maintainer email on the registry record is greyson@crewai.com. The package CrewAI’s own docs tell you to install is the Python distribution crewai 1.15.39.

The third is reading stars as the install. CrewAI led this trio by 15,977 stars at 11:45 UTC. The download window that ended November 8 had LangGraph ahead by a wide margin, and that margin is computed on the install page, not re-summed here. The live counters are the downloads tracker and the stars tracker.

Multi-agent orchestration is the research question of which calls to make together. How AI agents work is the loop these libraries implement. This page is the purchasing and lockfile decision in front of both.

Frequently asked questions

Which should you pick, LangGraph, CrewAI, or the OpenAI Agents SDK?

LangGraph, if the run has to resume from a checkpoint. CrewAI, if the work is roles and Flows on Python 3.10 through 3.13. The OpenAI Agents SDK, if you will host a small handoff loop and you want tracing on by default. The libraries are MIT. LangSmith, CrewAI AMP, and the model API are the bills.

Are LangGraph, CrewAI, and the OpenAI Agents SDK free?

The libraries are. LangSmith Developer is $0 for one seat and 5,000 base traces a month. Plus is $39 per seat and 10,000 base traces. CrewAI AMP Basic is free for 50 workflow executions a month. Enterprise is custom. The OpenAI SDK has no seat price. Traces in the OpenAI dashboard are described as free, and they are unavailable under Zero Data Retention.

What versions were current on November 10, 2026?

langgraph 1.2.20, crewai 1.15.39, and openai-agents 0.24.0 on PyPI. @langchain/langgraph 1.7.3 and @openai/agents 0.21.0 on npm. GitHub’s latest LangGraph tag was sdk==0.4.14.

Which model does the OpenAI Agents SDK use if you do not set one?

gpt-5.6-luna, with reasoning effort none and verbosity low. The models page points at gpt-5.6-sol for frontier work. CrewAI’s test command defaults to gpt-6-luna and OpenAI only. LangGraph does not name a default model.

Do all three support MCP?

CrewAI and the OpenAI Agents SDK document client transports in the framework docs, including stdio and streamable HTTP. LangGraph calls tools you pass into a node. LangChain’s MCP adapter is a beta LangChain feature. Agent Server can also publish a deployed graph at /mcp.

Can a LangGraph run resume after a crash?

With PostgresSaver or SqliteSaver, yes. With InMemorySaver, the checkpoint dies with the process. CrewAI’s @persist defaults to SQLite and can resume or fork a Flow. An OpenAI Agents SDK session stores the conversation and resumes an approval. It is not a graph checkpoint.

Does CrewAI 1.15.39 support Python 3.14?

The PyPI metadata says no. It requires Python at least 3.10 and below 3.14.

Is npm install crewai the CrewAI framework?

The npm package crewai was 1.0.1 from August 21, 2024, pointed at jaafarskafi1/crew-js. CrewAI 1.15.39 is the Python package.