Can I self-host a LangGraph app without a key?

[cf-sewe]

I am in the process of migrating my chat app from LangServe to LangGraph and want to self-host it in Docker (as before).
Is this deployment scenario still valid? I only have a free LangSmith account.

From documentation I think it will be possible to run langgraph build, but running the actual LangGraph Docker container seems to require an API key. What does LANGGRAPH_AUTH_TYPE=noop do in this context?

What options do I have?

[PlebeiusGaragicus]

Also looking for a solution of a fully self-hosted and self-managed solution for deploying my LangGraph agents - no LangSmith monitoring desired. Working on researching this right now - will report back.

In the past I had used FastAPI to create an endpoint in which I called graph.astream_events() directly. Unsure if a 'proper' LangGraph deployment will allow for free usage.

[PlebeiusGaragicus]

I love LangChain and follow everything they put out. But yes, I focus on fully self-hosted solutions.

They built something great and they deserve the profit they generate from it.

I recently discovered this: agent-service-toolkit - another instance of LangGraph being self-hosted with FastAPI. I've now seen a few examples of this and have done it myself. It's pretty straight-forward.

Don't want to pay... do it yourself ;)

As long as they allow that I won't be mad.

[kisanmajumder]

@PlebeiusGaragicus @vemonet Hey guys, thanks for sharing these awesome alternatives, I am also trying to find a open source solution to deploy langgraph application.
But I am not able to understand when we use fast API for the deployment, how will we incorporate redis for task queue and postgres for persistence in the API. While going through the documentation on langgraph self hosted deployment, I do see they are integrating redis and postgres.

Not a langgraph expert but If I choose the FAST api route for deployment, do I need redis and postgres too.
I want to know what will be the shortcomings in my chatbot if I don't integrate those

[kenberland]

I'm using FastAPI, as above, and I don't need Redis or Postgres.

[eyurtsev]

@vemonet thank you for the feedback!

I hope this: #1604 (comment) answers some of the concerns.

Open source self-hosted deployment without usage limitations, and no automatic logs sent to the proprietary company server is paramount when building sustainable systems for scientific research.

• You get 1M node executions per year with the Developer account (which is also free for sign up).
• Disable tracing LANGCHAIN_TRACING_V2=false.

Logging is mandatory, proprietary and external

• Disable tracing LANGCHAIN_TRACING_V2=false.

Actually deploying LangGraph self-hosted

Follow guidelines here: https://langchain-ai.github.io/langgraph/how-tos/deploy-self-hosted/

LangGraph Server has a different architecture from LangServe, and requires Redis and Postgres. These enable supporting a number of useful features out of the box:

• human-in-the-loop -- which involves interruption of a workflow for an indefinite amount of time before resuming it.
• memory -- with support for short term and long term memory.
• Additional features are listed here: https://langchain-ai.github.io/langgraph/concepts/langgraph_platform/#why-use-langgraph-platform

I'd recommend trying out spinning up a local sever with this: https://langchain-ai.github.io/langgraph/tutorials/langgraph-platform/local-server/ and testing things out!

[hinthornw]

Thanks for the feedback. I take responsibility for wherever we under- or miscommunicate. I also assume good intent by the original poster. I understand it can be frustrating if it's hard to find the free fast path to deploy and self-host things.

We developed the LangGraph API as an optional, opinionated way to host LangGraph workflows. There are many examples of people deploying LangGraph objects behind custom fastapi, starlette, django, and other web frameworks. This is great!

We also believe that some teams and enterprises benefit from a standardized set of APIs with built-in implementations for durability, autoscaling, auth, multitasking, etc. that you can get out of the box. This functionality has never been open source. Pricing is hard to get right; once it is generally available, we will provide more clear pricing.

The original post contains a few misunderstandings: I probably haven't been sufficiently clear in the docs.

• LangGraph OSS is a python library; almost the entirety of our docs so how to be running things locally; it's trivial to integrate that in your own API or existing codebase.
• The LangGraph API was always closed source; by offering a product we believe enterprises and some teams will find useful, we hope to guarantee longevity for LangGraph as a whole so we can invest more in making LangGraph OSS great.
• LangSmith does support self-hosting for enterprises, and we have a small team and need to balance our support commitments and so do so based on licenses today. Self-hosting in general comes with an assumption of support, and time is scarce.
• LangGraph OSS was explicitly designed to support cycles; It is not a DAG framework. This was in the first paragraph of the LangGraph intro at day 1.
• We did not kill LangServe; as a small team, we cannot commi to maintaining it. The community can continue to maintain it if they want.
• LangGraph API doesn't share any code from LangServe. Yes the API component is starlette (python) right now.

We are fully committed to supporting our open source work; the top 13 contributors of the OSS lib and vast majority of work done on the lib are all from our team. We spend a ton of time, thought, and energy into making things that you can all use for free. We also want to make something that can sustain our team so we can build even greater things for everyone. We will certainly make mistakes along the way, but we are always trying to do better.

I think because of some of the mis-understandings in the original post, I'm going to hide it (still visible if you expand) because I believe it misrepresents LangGraph and what options are available to users.

If you don't read any of the above:

• LangGraph OSS is an MIT-licensed library that provides an agent framework with durable execution. You can use it in any codebase and serve it however you see fit.
• LangGraph API is a proprietary runtime with built-ins, default routes, and a convenient development UI that we believe makes it faster to get into production and is especially useful for larger teams. We think it solves some (not all) of the harder problems about agent UX's.

[eyurtsev]

Self-hosted lite offers up to 1 million nodes executed per year for free. You can sign up for a Developer account on LangSmith. The Developer account is also free up to 5k trace executions per month.

How to guide: https://langchain-ai.github.io/langgraph/how-tos/deploy-self-hosted/

Up to date pricing information is here:

• https://www.langchain.com/pricing-langgraph-platform
• https://www.langchain.com/pricing-langsmith

---

Deploying LangGraph Server requires both PostgreSQL and Redis. LangGraph Server uses a web-queue-worker architecture that maintains the queue even during application crashes or horizontal scaling.

LangGraph Server provides a built-in persistence layer which can be used for storing application data (e.g., chat history), supproting human-in-the-loop workflows, and allows recovery after an exception occurs in the middle of a workflow (i.e., resuming without repeating previously done work).

LangGraph Server has a number of additional features that you'll likely want when deploying applications to production (e.g., implementation of double-texting strategy which allows you to handle a situation when a user sends two consecutive messages to an agent before it finished responding to the first one).

You can read more about the feature set here:

https://langchain-ai.github.io/langgraph/concepts/langgraph_platform/#why-use-langgraph-platform

[fhkingma]

For the self-hosted lite offering, does the container need an internet connection in order to validate the provided Langsmith key? I can't find that in the documentation unfortunately.

[hinthornw]

Yes for the lite offering. The air-gapped options would require a license key.

[PlebeiusGaragicus]

@kisanmajumder I think you can just self-host redis and postgres in their own Docker containers alongside your FastAPI LangGraph API. This is what's done in a development environment when you use LangGraph Studio.

Self-sovereignty demands fully self-hosted options. It looks like LangChain has very good Enterprise support - but even if less than 1% of users fall into this category, owning the full stack yourself is empowering.

[kisanmajumder]

@PlebeiusGaragicus Thanks for the insights

[hinthornw]

Yeah, you're always welcome to use FastAPI and any/all the open source building blocks we've built for you to run it your way :)

We do believe the functionality we've built into the platform is useful for a lot of folks, especially as teams grow, but I totally understand the desire to write all the server things yourself