Mastra vs LangChain vs LlamaIndex: TypeScript AI Agents (2026)

The ShipGarden gallery keeps a wall for the tools you actually build agents with, and this quarter three names kept coming up in the same breath: Mastra, LangChain.js, and LlamaIndex.TS. All three promise a TypeScript team the same thing: a way to wire a model to tools, memory, and your own data without hand-rolling the orchestration. They do not deliver it equally, and one of them quietly stopped delivering it at all this year.

We build with these before we shelve them, so this is the decision framed the way we actually make it: by what you are shipping and what language your team lives in, not by whose feature list is longest.

Quick answer (2026)

If you are building a TypeScript Next.js SaaS and want a framework designed for your language from the ground up, Mastra is the current pick. It is TypeScript-native, moves fast (@mastra/core v1.51.0 shipped July 15, 2026), and treats agents, workflows, memory, and retrieval as first-class parts of one package.

If you are already invested in the LangChain ecosystem, or you need the widest set of integrations plus a mature tracing and evaluation story, LangChain.js is the safe, MIT-licensed default.

LlamaIndex.TS is still strong at retrieval, but its TypeScript package was archived in 2026, so for a new TS project you are choosing between a frozen library and running LlamaIndex in Python. The full reasoning is below.

The three at a glance (2026)

Star, license, and version figures pulled from each project's GitHub repository in July 2026: Mastra, LangChain.js, and LlamaIndex.TS.

Mastra: built for TypeScript, moving fast

Mastra is the youngest of the three and the only one written as a TypeScript framework rather than ported into one. That shows up in the ergonomics: agents, tools, workflows, memory, and RAG live behind one typed API, and it deploys cleanly to Vercel, a Node server, or Cloudflare. At 26.2k stars with commits landing daily, it has the momentum of the moment, and its core is Apache-2.0 and free to self-host per its repository and docs (2026).

Where it loses: youth is a real cost. A fast release cadence means the API still moves under you, so pinning versions matters. The core is Apache-2.0, but a directory named ee/ in the repo is commercially licensed, so read the boundaries if enterprise features tempt you. Its integration catalog is narrower than LangChain's, and the managed Mastra Cloud is a separate paid product, not part of the framework.

LangChain.js: the mature default

LangChain.js is the framework you reach for when the answer is "there is already an integration for that." It carries the largest connector catalog of the three, ships under a permissive MIT license, and pairs with LangGraph for stateful multi-step agents and LangSmith for tracing and evaluation. At 17.9k stars on @langchain/core v1.2.3 (July 2026), it has the deepest community and the most Stack Overflow answers when you are stuck, all documented at its repo and JS docs.

Where it loses: the JavaScript library began as a port of the Python project and has historically trailed it on new features. The abstractions can feel heavy for a small agent, where you write more framework than logic. And the parts that make it shine in production, LangSmith observability in particular, are a separate commercial layer, so "free framework" and "free to operate well" are not the same sentence.

LlamaIndex.TS: the archived specialist

This is the honest one. LlamaIndex is the best-known name in retrieval-augmented generation, and the Python project is thriving: run-llama/llama_index sits at 50.9k stars under MIT and ships constantly (2026). The TypeScript port is a different story. As of 2026 the LlamaIndexTS repository is archived and read-only, with its last release in December 2025. For a pure-TypeScript team that means no new features and no security fixes from here.

It still installs and handles straightforward retrieval, so it is not useless. But building a new 2026 project on an archived dependency is a decision you should make with eyes open, not by accident. If RAG is your core and you can run a Python service, the maintained Python LlamaIndex is the stronger home. If you must stay in TypeScript, Mastra and LangChain.js now both cover retrieval well enough that reaching for a frozen library is hard to justify.

How to actually choose

• Greenfield TypeScript Next.js SaaS, agents plus workflows: start with Mastra. The TS-native ergonomics save the most time when nothing legacy is pinning you.
• You need a specific integration, or you already run LangChain, and you want evals and tracing: LangChain.js, with LangSmith if you are willing to pay for observability.
• RAG is the whole product and Python is on the table: the Python LlamaIndex, not the archived TS port. Otherwise use Mastra's retrieval or LangChain.js.
• Genuinely unsure: build the same three-step agent in Mastra and LangChain.js over one afternoon. The SDK that fits your head, not the one with the most stars, is the right answer.

Cost, honestly (2026)

All three frameworks are free and open source; your real bill is model API calls plus wherever the agent runs. The money question is the managed layer sitting next to each one: Mastra Cloud, LangChain's LangSmith and LangGraph Platform, and LlamaCloud are separate commercial products with their own tiers, and those tiers change often, so price them on each vendor's current pricing page before you commit. None of them is required to ship. You can run any of these frameworks entirely self-hosted and pay only your model provider.

We test these building blocks before we recommend them. The Mastra starter template review walks the three deploy paths for a Mastra agent hands-on, and the LlamaIndex.TS RAG starter review shows what the retrieval path actually looks like if you still want it.

FAQ

Your move: pick the two frameworks that fit your stack, build the same small agent in each this week, and let the developer experience, not the star count, make the call.