AutoGPT vs OpenHands vs MetaGPT vs CrewAI: Open-Source AI Agent Frameworks (2026)

These are the four most-starred open-source AI agent frameworks in our AI directory, and by mid-2026 they have almost nothing in common beyond the word "agent". One has pivoted into a paid hosted platform with its original CLI formally abandoned. One is mid-rebrand, moving code between repos while its terminal UI counts down to a sunset date. One has not shipped a release in sixteen months. One ships multiple releases a week.

That spread is the honest thesis of this comparison: GitHub stars measure accumulated attention, not current momentum. AutoGPT's 185.6k stars are more than double second-placed OpenHands' 80.9k, yet most of those stars were earned by a CLI experiment its own README now calls unsupported. Meanwhile the smallest project of the four, CrewAI, has the fastest release cadence and the cleanest license.

Everything below is sourced from each project's own repository, docs, and pricing pages, checked on 21 July 2026. Where a number is a project's claim about itself, we label it self-reported.

AutoGPT: the 2023 phenomenon is now a paid platform

AutoGPT (★185.6k ⑂46.1k, RECATOOLS traction data, 21 Jul 2026) earned its stars as the 2023 autonomous-agent experiment everyone ran once. That code still exists — under classic/ — but the project's own README for it is blunt: "This project is unsupported, and dependencies will not be updated," recommending the codebase "for educational purposes only." If a tutorial tells you to run the AutoGPT CLI in 2026, it is describing a museum piece.

What Significant Gravitas builds today is the AutoGPT Platform: a visual, no-code agent builder where workflows are described "in plain English" or assembled step by step, then run on demand, on a schedule, or from a trigger. The hosted Platform is publicly available with no waitlist, and it is a paid service with usage-based agent runs: Pro at $42.50/month and Max at $272.00/month, both billed annually (agpt.co/pricing, 21 Jul 2026), plus a pay-as-you-go credit wallet. There is no free hosted tier — the README's own words are "Self-hosting is the free path," via docker compose up -d --build inside autogpt_platform/ with your own model API keys.

The license is the detail most coverage gets wrong. The repo is split-licensed: everything inside autogpt_platform/ is under the PolyForm Shield License 1.0.0 — per the project's own licensing table, free for personal and internal business use but it "cannot be sold as a competing hosted service" — while everything outside that folder, including the classic agent, Forge, and AG Benchmark, remains MIT. GitHub's API reports the composite file as NOASSERTION, which is your machine-readable hint that this is not plain MIT.

Activity is healthy: platform releases land roughly weekly, the latest being v0.6.68 on 17 July 2026 — though release tags still carry a -beta- marker despite the hosted product being marketed as generally available. No current benchmarks are self-published for the platform.

OpenHands: strong coding agent, mid-rebrand turbulence

OpenHands (★80.9k ⑂10.3k) began as OpenDevin, became All Hands AI's flagship, and has since renamed its whole GitHub org from All-Hands-AI to OpenHands. The README now positions it as "the self-hosted developer control center for coding agents and automations" — able to run its own agent or drive Claude Code, Codex, Gemini, and other ACP-compatible agents across local, remote, and cloud backends.

The 2026 transition is real and worth understanding before you adopt it. Agent Canvas, the new browser-based control center, launched to the community in beta on 3 June 2026 (npm @openhands/agent-canvas, or a one-line Docker run). The Software Agent SDK is the officially supported core going forward — Python, plain MIT, very active (v1.36.1 on 15 July 2026). The legacy local GUI is moving to a separate repo in maintenance mode, and the CLI/TUI sunsets on 1 October 2026, with headless/ACP modes supported indefinitely. The team's own transition FAQ admits Agent Canvas still needs "work to do around stability."

Licensing mirrors AutoGPT's pattern in miniature: the flagship repo is MIT except the enterprise/ directory, which sits under the PolyForm Free Trial License 1.0.0. The project itself calls this "the annoying dual license" and says the mostly-enterprise-licensed backend will be moved out of the flagship repo. The successor SDK repo is plain MIT with no carve-outs.

On capability: the SDK README displays a SWE-Bench score of 77.6 — a self-reported figure, shown as a badge linking to the team's methodology sheet, with the subset and date not stated in the README itself. Treat it as OpenHands' own claim, not an independent result. MCP support is built into the SDK (stdio and URL servers, tool filtering, automatic OAuth) with one documented caveat: OAuth MCP servers need an interactive browser step, so they are "not suitable for fully automated/headless workflows." OpenHands Cloud offers a free Individual tier capped at 10 conversations per day, with LLM access "at cost, with no markup"; Enterprise is custom-priced, self-hosted in your own VPC via Kubernetes if you want. The company announced an $18.8M Series A in November 2025 (its own blog), led by Madrona.

MetaGPT: the research star whose OSS repo has gone quiet

MetaGPT (★69.4k ⑂8.8k) is the "AI software company" framework: give it a one-line requirement and its simulated product managers, architects, and engineers produce user stories, designs, APIs, and code, following the philosophy Code = SOP(Team). The academic pedigree is genuine — the MetaGPT paper was an ICLR 2024 oral, and the team's AFlow follow-up an ICLR 2025 oral.

The repository state in 2026, stated with dates rather than adjectives: the last release is v0.8.2, published 9 March 2025 — sixteen months before this guide — and the last push to main was 21 January 2026, a Windows-terminal PR merge, with roughly four commits since mid-2025. Six months with zero pushes as of 21 July 2026. Meanwhile the company's commercial product MGX (mgx.dev) now 301-redirects to atoms.dev, Foundation Agents' AI app-building platform, whose own footer says it is built on open-source technology including MetaGPT and OpenManus. The repo's homepage field points at atoms.dev too. The energy has visibly moved.

The license is the simple part: plain MIT, no carve-outs. Install is pip install --upgrade metagpt, with one dated constraint straight from the README: Python "3.9 or later, but less than 3.12" — awkward on 2026 default toolchains. Our GitHub code search found no MCP support (zero path hits). The benchmark you will see quoted — 85.9% Pass@1 on HumanEval (and 87.7% on MBPP) — is self-reported, from the 2023–24 paper, with GPT-4 as the backend. It was a strong result in its era; it is not comparable to 2026 coding-agent benchmarks and says nothing about current capability.

CrewAI: smallest star count, fastest pulse

CrewAI (★55.6k ⑂7.8k) is an open-source Python framework for "production-ready multi-agent workflows" built on two primitives: Crews — role-based teams of autonomous agents that collaborate — and Flows — event-driven workflows for precise, deterministic control. The old "built on LangChain" framing is obsolete; the README describes a purpose-built, lightweight Python core.

It is the activity leader of this comparison by a wide margin. The latest release, 1.15.5, was published 20 July 2026 — the day before our data pull — and the project ships multiple releases per week. The license is the cleanest of the four: plain MIT, no carve-outs anywhere in the OSS repo. Install is uv pip install crewai (Python ≥3.10 <3.14), with a 'crewai[tools]' extra.

MCP is first-class, with a documented boundary. Agents take a simple mcps field, or you can use MCPServerAdapter from crewai-tools, over stdio, SSE, or streamable HTTP transports. The docs are explicit that only MCP tools are integrated — "Other MCP primitives like prompts or resources are not directly integrated" — and they carry an SSE DNS-rebinding warning plus a reminder to only connect servers you trust.

The commercial layer is the CrewAI AMP Suite: a Free tier with 50 workflow executions per month (visual editor, AI copilot, GitHub integration) and a custom-priced Enterprise tier with CrewAI-hosted or private infrastructure. No dollar figures are published for Enterprise, and crewai.com/enterprise itself returned a 404 on 21 July 2026 — the offer is presented via the pricing page. CrewAI publishes no coding benchmarks; its headline numbers are marketing metrics from its October 2024 $18M funding announcement, which we treat as self-reported and vintage.

License accuracy: two clean MITs, two split repos

"Open source" hides real differences here, and half the coverage of these projects mis-states them. Two of the four repos are plain MIT; two are split-licensed with a PolyForm-licensed directory that changes what you can build commercially. The nuance matters most if you plan to host any of these as a service for others.

Practical translation: the MIT-covered code in all four is free to self-host for any internal use, and so is AutoGPT's PolyForm Shield platform — its one line is that you cannot sell it as a competing hosted service. OpenHands' enterprise/ directory is the stricter case: PolyForm Free Trial is a trial licence, and the docs state an enterprise licence is required beyond one month, so treat that directory as commercial code rather than open source. Self-hosting OpenHands itself does not require it.

Feature-by-feature

Which one should you pick?

Coding tasks against real repositories → OpenHands. It is the only one of the four positioned as a coding agent ("control center for coding agents"), the only one with a current benchmark claim (SWE-Bench 77.6, self-reported), and its MIT SDK plus ACP support means it can also drive other coding agents. Go in with eyes open about the transition: Agent Canvas is beta, the TUI dies on 1 October 2026, and the flagship repo's license is split until the enterprise backend moves out.

Role-based multi-agent business workflows in Python → CrewAI. Crews and Flows map cleanly onto "a team of specialists plus an orchestrator", the license is clean MIT, MCP is first-class (tools-only), and the release cadence — 1.15.5 shipped the day before we pulled data — is the strongest maintenance signal in this comparison. The trade-off is an opaque Enterprise price and no published coding benchmarks.

Visual agent platform with a hosted option → AutoGPT Platform. If you want triggers, schedules, and a no-code builder rather than a Python codebase, AutoGPT is the only one of the four built around that model, with flat hosted plans and a free self-host path. Two caveats: the platform folder's PolyForm Shield license means you cannot resell it as a hosted service, and none of this applies to the classic CLI, which is formally unsupported.

MetaGPT → historical and research interest. Its software-company simulation remains a genuinely influential idea, and the MIT license means the code is yours to study and extend. But adopting it for new production work in 2026 means building on a repo whose last release was 9 March 2025 and whose last push was 21 January 2026, with the company's effort going to Atoms. Those are the facts; they do not diminish what the project contributed, but they should decide your risk calculus.

No "best overall" award here, because the four are not playing the same game: a no-code automation platform, a coding agent, a research framework, and a Python orchestration library only compete at the edges. If your use case is workflow automation rather than autonomous agents, our n8n vs Dify vs Langflow vs Flowise comparison is the better shortlist.

The ASEAN angle: self-hosting is the real story

For teams in Singapore, Indonesia, and the wider region, the practical headline is data control. All four are self-hostable at zero license cost for internal use — including AutoGPT's PolyForm-covered platform — which means prompts, agent traffic, and repository contents can stay on infrastructure you control, a meaningful posture under Singapore's PDPA or Indonesia's PDP Law. If you have self-hosted anything with Docker before (our OpenClaw self-hosting walkthrough covers the pattern), AutoGPT's and OpenHands' compose/run paths will feel familiar. On the managed side, be realistic: none of the four documents an ASEAN or APAC cloud region, and none publishes data-residency options on the pages we reviewed. The closest substitutes are deploy-anywhere claims — OpenHands Enterprise in your own VPC via Kubernetes, CrewAI AMP's on-premise option. All four companies are outside the region, so support timezones are a fair practical consideration, noted as absence of evidence rather than a known problem.

FAQ

Are these actually open source?

All four publish their code, but only MetaGPT and CrewAI are plain MIT throughout. AutoGPT puts autogpt_platform/ under PolyForm Shield 1.0.0 (free to use, cannot be resold as a competing hosted service), and OpenHands puts enterprise/ under PolyForm Free Trial 1.0.0, a trial licence requiring a paid licence beyond one month. For internal use, all four are free so long as you stay outside that OpenHands enterprise/ directory; for building a commercial hosted service on top, read the split licenses first.

Do I need an agent framework at all, or is an LLM API enough?

If your task is a single prompt-and-response — summarise, classify, draft — a direct LLM API call is simpler and cheaper. These frameworks earn their complexity when work spans multiple steps, tools, or roles: CrewAI coordinates specialist agents, OpenHands drives a coding agent through a real repository, AutoGPT runs workflows on schedules and triggers. Start with the API; adopt a framework when orchestration, not generation, becomes your problem.

Can they run fully local?

All four are self-hostable and all four let you bring your own model keys: AutoGPT's self-host path requires your own keys, OpenHands advertises "use with any LLM", MetaGPT configures models via config2.yaml, and CrewAI is provider-agnostic. Whether the model itself runs locally depends on pointing that configuration at a local endpoint — and on whether a local model is strong enough for agent workloads, which none of the four guarantees.

What is the cheapest way to try each?

AutoGPT: the no-signup instant demo, or the free self-host path via Docker Compose with your own keys. OpenHands: the free Cloud Individual tier (10 conversations/day) or npm i -g @openhands/agent-canvas. MetaGPT: pip install --upgrade metagpt plus your own model key — mind the Python <3.12 ceiling. CrewAI: uv pip install crewai, or the AMP Free tier's 50 workflow executions per month. In every case the real cost is model usage, not the framework.