7 Open Source AI Agents Worth Building On in 2026

7 Open Source AI Agents Worth Building On in 2026

Anyone comparing open source AI agents this year runs into the same wall: GitHub stars tell you nothing about the license terms, and every project calls itself "the leading framework." The seven below are the ones with real, dated activity in 2026 and a license you can actually read before you build on them.

Open source AI agents are agent frameworks and runtimes whose code you can inspect, self-host and modify, released under licenses ranging from permissive MIT and Apache 2.0 to restricted "fair-code" terms that limit resale or multi-tenant hosting. The right pick depends on what you're building and who else will run your code.

Comparing open source AI agents

LangChain and LangGraph

LangChain is an MIT-licensed open source framework with a prebuilt agent architecture and integrations for any model or tool provider, according to its own FAQ. Its create_agent function gives you a ReAct pattern running on LangGraph's durable runtime. LangGraph itself is the lower-level piece: a durable execution runtime focused on agent orchestration, with persistence, human-in-the-loop interrupts, streaming and checkpointing built in. LangChain's own documentation lists Klarna, Uber and J.P. Morgan among companies using LangGraph in production. Installation is a single pip install -U langgraph command, and LangChain recommends using its higher-level agent abstractions unless you specifically need LangGraph's fine-grained control over mixing deterministic and model-driven steps in one graph.

CrewAI

CrewAI is an open source Python framework, licensed under MIT, organized around two units: Crews, teams of role-based agents that collaborate autonomously, and Flows, event-driven workflows that manage state and control execution precisely. Its documentation reports more than 100,000 developers have completed its community certification courses. CrewAI ships frequent point releases: version 1.15.22, released 16 September 2026, added an llm_overlay context variable that routes individual agent roles to different models, human-feedback capture inside tracing, and OpenRouter as a supported embedding provider. The prior release, 1.15.21 on 9 September 2026, added telemetry for checkpoint runtime and CLI usage. CrewAI has first-class support for MCP as a way to attach tools, plus custom MCP servers and native sandbox tools for E2B and Daytona, and it supports the Agent-to-Agent (A2A) protocol for connecting separate crews.

OpenHands

OpenHands is MIT licensed and split into several components documented in its own repository map: Agent Canvas is the open source browser client and control center for agent conversations; the Software Agent SDK is a composable Python library for building agents that work with code, packaged together with an Agent Server that exposes execution, conversations and workspaces over REST and WebSocket APIs; a separate Automation Server handles scheduled and event-driven automation; and a Sandbox Server is the community-supported standalone API and control plane for sandboxed environments. OpenHands' own use-case guide points the project at vulnerability remediation, automated code review, QA testing, incident triage, COBOL modernization and dependency upgrades, rather than general-purpose chat.

goose

goose is Block's native, open source AI agent, shipped as a desktop app for macOS, Linux and Windows, a full CLI, and an API, built in Rust. It runs on 15 or more model providers, including Anthropic, OpenAI, Google, Ollama, Azure and Bedrock, and it can use your existing Claude, ChatGPT or Gemini subscription through the Agent Client Protocol (ACP) instead of a separate API key. goose's own site documents 45,000-plus GitHub stars, 500-plus contributors and 70-plus MCP extensions. On 7 April 2026, goose announced it had moved to the Agentic AI Foundation (AAIF), a Linux Foundation project, to keep the project vendor-neutral and community-governed rather than controlled by a single company. It is licensed under Apache 2.0 and also supports subagents for parallel tasks, portable YAML "recipes" for sharing workflows, and MCP Apps that render interactive UI inside the desktop client.

AutoGPT

AutoGPT, from Significant-Gravitas, was one of the first widely used autonomous agents and its GitHub repository now carries two licenses at once. Code inside the autogpt_platform folder, the newer visual agent-building platform, is under the PolyForm Shield License, which permits any use except building a product that competes with AutoGPT or its affiliates. Everything outside that folder, including the original standalone AutoGPT Agent, the Forge framework and the AG Benchmark, is MIT licensed with no such restriction. That split matters if you plan to build a commercial product on top of the platform code specifically rather than the classic agent.

Dify

Dify, from LangGenius, is a platform for building agentic workflows and RAG pipelines with a visual, collaborative workspace, deployable on cloud, VPC or self-hosted infrastructure, and it is one of the tools indexed on AI Agents Listing. Its license is a modified Apache License 2.0: you can use Dify commercially, including as a backend service or application development platform, but running it to operate a multi-tenant service, where one tenant equals one workspace, requires a separate commercial license from LangGenius unless explicitly authorized in writing. The license also blocks removing or modifying the Dify logo and copyright notices in its console or applications. Outside those two conditions, Apache 2.0 terms apply.

n8n

n8n is a workflow automation platform with native AI capabilities, more than 400 integrations, and support for building AI agents alongside deterministic, rule-based steps in the same workflow. It is not released under an OSI-approved open source license: it uses the Sustainable Use License, a "fair-code" license n8n created in 2022 after switching away from Apache 2.0 with Commons Clause on 17 March 2026. The Sustainable Use License lets you use, modify and redistribute the software for free, but only for internal business purposes or non-commercial use; distributing it to others, or hosting it as a paid multi-tenant service, needs a separate commercial agreement. Files with .ee. in their name fall under the separate n8n Enterprise License instead. n8n's AI agent use case documentation covers routing support tickets, RAG over internal documents, and MCP-connected lead discovery.

How to choose

Start with the license, not the star count. LangChain, LangGraph, CrewAI, OpenHands and goose are permissively licensed (MIT or Apache 2.0), so you can fork, resell or embed them without asking anyone. Dify, n8n, and the platform half of AutoGPT carry commercial restrictions that only bite if you plan to host the software for other tenants or compete directly with the vendor; for internal use they behave like open source software. After the license, match the shape of the tool to the job: LangGraph and CrewAI are code-first frameworks for developers who want to define agent logic directly; OpenHands and goose are agent runtimes you point at a task rather than a library you import; Dify and n8n are visual builders aimed at teams that want non-engineers editing the workflow too.

What's new (as of 23 September 2026)

• CrewAI shipped v1.15.22 on 16 September 2026, adding an llm_overlay context variable to route individual agent roles to different models, human-feedback capture inside tracing, and OpenRouter as a supported embedding provider.
• CrewAI's previous release, v1.15.21 on 9 September 2026, added telemetry to track checkpoint runtime and CLI usage.
• goose announced its move to the Agentic AI Foundation, a Linux Foundation project, on 7 April 2026, changing its governance from a single-vendor project to a vendor-neutral foundation.

Key takeaways

• Open source AI agent frameworks split into two license groups: MIT or Apache 2.0 tools (LangChain, CrewAI, OpenHands, goose) with no commercial-use restriction, and fair-code or modified-Apache tools (n8n, Dify, the AutoGPT platform folder) that restrict reselling or multi-tenant hosting.
• CrewAI ships near-weekly point releases; the most recent, v1.15.22 on 16 September 2026, added per-role model routing through a new llm_overlay variable.
• goose left single-vendor governance at Block for the Linux Foundation's Agentic AI Foundation on 7 April 2026 and documents 70-plus MCP extensions.
• AutoGPT's repository carries two different licenses: MIT for the original agent and Forge, PolyForm Shield (noncompete) for the newer AutoGPT Platform folder.
• Dify's license permits commercial use but blocks running it as a paid multi-tenant service without a separate agreement from LangGenius.

FAQ

Is LangChain open source?

Yes. LangChain states in its own FAQ that it is an MIT-licensed open source library, free to use, with no restrictions on commercial use.

What license does n8n use?

n8n uses its own Sustainable Use License plus the n8n Enterprise License for .ee. files, both based on the "fair-code" model. It permits free use, modification and redistribution for internal business or non-commercial purposes, but hosting n8n as a paid service for other tenants requires a separate commercial agreement.

Can I use AutoGPT commercially?

It depends which part of the repository you use. Code outside the autogpt_platform folder, including the original AutoGPT Agent and Forge, is MIT licensed with no commercial restriction. Code inside autogpt_platform is under the PolyForm Shield License, which allows commercial use except building a product that competes with AutoGPT.

Does CrewAI support MCP servers?

Yes. CrewAI's own documentation describes first-class support for MCP servers as tools, including custom MCP servers over stdio, SSE and streamable HTTP transports, alongside native sandbox tools for E2B and Daytona.

Compare these frameworks against the hosted agents and MCP servers already reviewed on AI Agents Listing.