Best Open-Source AI Agent Frameworks (2026)

The most-starred open-source AI agent frameworks in mid-2026 are CrewAI (55,260 GitHub stars), LlamaIndex (50,751), SWE-agent (19,757), and Pydantic AI (18,300), all MIT-licensed, model-agnostic, and usable against OpenAI-compatible Chinese LLM endpoints. That last property is the reason this cluster lives on china-llm.com. None of these projects is tied to a single provider, so a team can point any of them at DeepSeek, Qwen, GLM, MiniMax, or Kimi with a base-URL and API-key change instead of a rewrite.

This hub is the map for the whole open-source-agent ecosystem as we found it in the last 30 days, mid-June to mid-July 2026. Below are two first-hand evidence blocks (a dated GitHub snapshot of the four leading frameworks and a measured OpenRouter run across five new open-weight models), an original pick-a-framework-by-job matrix that no single sub-article states, and a navigation table into all ten deep-dives.

The open-source agent cluster (verified 2026-07)

Each article below answers one question a team building agents on Chinese LLMs actually arrives with, from "what is this framework" to "which model do I pair it with."

The framework snapshot: what we pulled from GitHub on 2026-07-10

Star counts and release tags move daily, so we treat this as a dated snapshot rather than a standing fact. We read all four repositories from the GitHub public REST API on the same day to keep them comparable.

When we pulled the four repositories from the GitHub API on 2026-07-10, CrewAI led at 55,260 stars with its 1.15.2 release shipped two days earlier, LlamaIndex held 50,751, SWE-agent 19,757, and Pydantic AI 18,300. Every one carries an MIT license and ships in Python.

According to CrewAI's GitHub repository, the project is MIT-licensed and had shipped release 1.15.2 by our check, which lines up with its position as the most-starred and fastest-moving of the four. CrewAI is licensed under MIT and written in Python, a permissive footing that matters when you embed it in a commercial product.

Release cadence, not just star count, tells you where the energy is. According to LlamaIndex's GitHub repository, the data framework tagged v0.14.23 on 2026-06-24, a rhythm of frequent point releases that fits its role as the RAG workhorse of the set. SWE-agent is the outlier on cadence: its v1.1.0 has stood since May 2025, so its 19,757 stars reward a stable research tool rather than a weekly shipper.

What the frameworks actually do (and where they diverge)

CrewAI is a multi-agent orchestration framework that models work as role-playing crews of agents, while LlamaIndex is a data framework that specializes in retrieval-augmented generation over private documents. Both are model-agnostic, so either can drive a Chinese LLM through an OpenAI-compatible endpoint without code changes.

SWE-agent is a coding agent from Princeton that gives a model an Agent-Computer Interface for editing and running code, and it has held its v1.1.0 release since May 2025. Pydantic AI is a type-safe agent framework whose v2.8.0 shipped the same day we checked. The practical upshot: these four rarely compete head to head. You pick by the job, which is what the matrix below does.

Pick a framework by the job

This is the cross-topic synthesis the individual articles do not state, because each covers one project. Match the framework to the shape of the work, then read its deep-dive.

If you are still choosing between the two heavyweights for a mixed workload, the CrewAI vs LlamaIndex comparison walks the tradeoff in full. The short version: orchestration-first work leans CrewAI, retrieval-first work leans LlamaIndex, and both will happily run a Chinese backend.

The models teams pair them with: measured via OpenRouter

A framework is only half the stack. The other half is the model, and the last month brought a wave of new open-weight releases. We ran a small identical prompt against five of them via OpenRouter to capture real token counts, billed cost, and latency. These are routed measurements, so they need native re-verification against each provider's own endpoint before you budget against them.

When we called Qwen3-235B-A22B through OpenRouter on 2026-07-10, a 36-token prompt returned 66 tokens in 2.87 seconds and billed $0.00004485, the cheapest run in our set. GLM-5 was the slowest and priciest, taking 7.62 seconds and $0.00050005 for a 150-token completion.

The five open-weight models we measured spanned a wide cost band: Qwen3-235B-A22B billed $0.00004485 for its run, DeepSeek-V3.2 $0.00013608, MiniMax-M2.5 $0.0001914, and GLM-5 $0.00050005. For an agent loop that fires thousands of calls, that eleven-fold spread compounds into a real monthly difference.

Qwen3-235B-A22B is released under Apache 2.0. According to the Apache License 2.0, that grant permits commercial use, modification, and redistribution, which is why an Apache-licensed model plus an MIT-licensed framework is the least encumbered agent stack a team can assemble. GLM-5 is Zhipu's open release, and its $1.92 output rate is roughly nineteen times Qwen3-235B-A22B's $0.10, so on output-heavy agent runs the model choice moves the bill more than the framework does.

Two provider notes for budgeting. According to DeepSeek's API documentation, DeepSeek exposes an OpenAI-compatible surface, which is exactly what lets these frameworks target it without a custom adapter. According to MiniMax's open platform, MiniMax distributes its M-series models through a first-party API, so a same-model native call is the reference point our OpenRouter figure should be re-verified against.

One coding model sits outside the measured set on purpose. Kimi-Dev-72B, Moonshot's open coding model, is not listed on OpenRouter, so we did not call it and quote no first-hand numbers for it. Its widely cited ~60.4% SWE-bench Verified score is sourced research, not our measurement, and the best open-source coding model 2026 article treats it as such.

Methodology

Framework facts are live GitHub REST reads (stars, license, latest release, push date) captured on 2026-07-10; star counts and dates move, so we frame them as a dated snapshot. Model prices and latencies are OpenRouter-routed measurements from the same day and are flagged for native re-verification against each provider. The aggregator neither hosts nor resells any model or framework; we publish neutral editorial comparisons and refresh the numbers when a repository or price page changes. The full run is in the cluster evidence pack.

FAQ

What are the best open-source AI agent frameworks in 2026? By GitHub stars on 2026-07-10, the leaders are CrewAI (55,260), LlamaIndex (50,751), SWE-agent (19,757), and Pydantic AI (18,300). All four are MIT-licensed and model-agnostic, so the "best" one depends on the job: CrewAI for multi-agent teams, LlamaIndex for RAG, SWE-agent for coding, Pydantic AI for type safety.

Do these frameworks work with Chinese LLM APIs? Yes. Every framework in this cluster is model-agnostic and speaks the OpenAI-compatible protocol, so pointing one at DeepSeek-V3.2, Qwen3-235B-A22B, GLM-5, or MiniMax-M2.5 is a base-URL and API-key change. See the Chinese-LLM framework guide.

Which framework should I use for a coding agent? SWE-agent is purpose-built for it, giving the model an Agent-Computer Interface to edit and run code. For the model behind it, compare the open coding models in the best open-source coding model 2026 piece, where Kimi-Dev-72B's research SWE-bench score sits alongside Qwen3-Coder-Next.

How much do the paired open-weight models cost to run? In our measured runs, a short prompt cost $0.00004485 on Qwen3-235B-A22B, $0.00013608 on DeepSeek-V3.2, $0.0001914 on MiniMax-M2.5, and $0.00050005 on GLM-5. Output rate drives the spread: Qwen3-235B-A22B outputs at $0.10/M versus GLM-5's $1.92/M.

Are these frameworks free to use commercially? All four (CrewAI, LlamaIndex, SWE-agent, Pydantic AI) are MIT-licensed, which permits commercial use. Pair one with an Apache 2.0 model such as Qwen3-235B-A22B and the whole stack is permissively licensed for production.

Why measure via OpenRouter instead of each provider directly? OpenRouter gives one consistent harness to compare five models on the same prompt in a single run. The tradeoff is that prices and latency are routed, not native, so we flag every number for re-verification against the provider's own endpoint before you commit a budget.

Author: Kevin Fan, Customer Success Manager at China LLM Directory, specializing in Chinese LLM ecosystem pricing, AI infrastructure benchmarking, and model evaluation. Last verified: 2026-07-10.