Choose an AutoGPT alternative on lifecycle first: AutoGPT Classic is explicitly unsupported and ships with known dependency vulnerabilities, while the AutoGPT Platform in the same repository published v0.8.0 on September 19, 2026 and is still under active development. They are two products under one name and two licences, so "is AutoGPT dead" has no single answer — and four of the projects that alternatives lists routinely put beside it are themselves archived, in maintenance mode, or untouched since 2025.
The repository Significant-Gravitas/AutoGPT is not archived. The GitHub API returned archived: false, disabled: false, 187,475 stars and a push dated September 21, 2026, with agpt.co as its homepage (checked September 21, 2026; the star count and push timestamp move). What has concluded is the half in classic/.
AutoGPT Classic is unsupported — the Platform is not
The Classic README states it plainly: "This project is unsupported, and dependencies will not be updated. It was an experiment that has concluded its initial research phase." It also warns that the codebase "has known vulnerabilities and issues with its dependencies" and should be used "for educational purposes only". The Classic setup documentation nevertheless still resolves, which is how readers keep starting an install nobody maintains. A page being online is not a statement of support.
AutoGPT Classic (classic/) | AutoGPT Platform (autogpt_platform/) | |
|---|---|---|
| Status | Unsupported; dependencies frozen | Active; autogpt-platform-beta-v0.8.0 published September 19, 2026 |
| Licence | MIT | PolyForm Shield License 1.0.0 — "Any purpose is a permitted purpose, except for providing any product that competes with the software" |
| Shape | Terminal loop configured by a .env file | Visual block graphs, plus an AutoPilot chat agent |
| Custom OpenAI base URL | OPENAI_API_BASE_URL in .env.template | AutoPilot only, self-hosted only; nothing in the block layer |
| Hosted option | None | agpt.co Pro and Max |
The licence split is worth reading twice, because the retired half is the permissive one. The repository's root LICENSE says: "Everything inside the autogpt_platform folder is under the Polyform Shield License. Everything outside the autogpt_platform folder is under the MIT License." A fork of Classic carries only MIT's attribution condition. The Platform's own LICENSE.md is PolyForm Shield License 1.0.0, and its Noncompete clause is wider than a ban on reselling the hosted product: "Any purpose is a permitted purpose, except for providing any product that competes with the software or any product the licensor or any of its affiliates provides using the software", with a Competition clause adding that goods and services "compete even when provided free of charge". Ordinary commercial and internal use is permitted; building something that competes is not. If the reason you are shopping is licensing rather than maintenance, that is the clause that decides it.
Why people leave, and where each reason lands
| What you relied on in Classic | Where it lands |
|---|---|
| A loop you could read and patch | OpenHands and Agent Zero are both MIT code you run and modify yourself; the AutoGPT Platform puts the agent in a visual block graph instead |
OPENAI_API_BASE_URL pointed at your own gateway | The narrowest loss. See the boundary section below — the Platform has no block-layer equivalent |
| Agents written as code you version-control | CrewAI or an equivalent Python framework; graph builders move that logic into a UI |
| MIT terms, because you ship a product on top | OpenHands, Agent Zero and CrewAI are MIT; the AutoGPT Platform is not |
| Nothing — you just want it maintained | The AutoGPT Platform is the vendor's own answer, and it is genuinely active |
The custom-endpoint boundary, stated exactly
This is the migration detail most alternatives pages skip, and it is the one that decides whether a gateway customer can stay. AutoGPT Classic's .env.template documents OPENAI_API_BASE_URL as a "Custom url for the OpenAI API, useful for connecting to custom backends". The AutoGPT Platform has no equivalent in its block layer.
Where a custom endpoint does work is the AutoPilot chat on a self-hosted install. The Platform's own documentation describes routing it "without an Anthropic, OpenAI, or OpenRouter key" through "any OpenAI-compatible HTTP endpoint you control", and its scenario table includes "A managed OpenAI-compatible API you don't pay AutoGPT for" pointed at its /v1 URL — so a remote managed gateway is explicitly in scope, not just localhost. The switches are CHAT_USE_LOCAL, CHAT_BASE_URL and CHAT_API_KEY in autogpt_platform/backend/.env. Four constraints come with them, read on September 21, 2026 from that documentation page and from the chat config it describes, backend/copilot/config.py:
| Constraint | What the source says |
|---|---|
| Self-host only | "the cloud agpt.co deployment routes AutoPilot through Anthropic / OpenRouter and ignores the variables below" |
| Bare model ids | "Bare model names ONLY — provider/model slugs (e.g. anthropic/claude-...) are passed through verbatim" |
| No extended thinking | The docs mark the Claude Agent SDK path unsupported under the local transport — "auto-downgrades to fast", and "the downgrade is logged at WARNING". config.py gives the reason: "the Claude Agent SDK CLI speaks Anthropic's wire protocol and cannot route to Ollama" |
| No key fallback | CHAT_API_KEY must be set explicitly; the transport deliberately will not fall back to another key already in the environment |
The bare-name rule is the one that quietly breaks gateways. Kunavo's catalog ids are unprefixed — claude-sonnet-4-6, claude-haiku-4-5 — so they match the format that documentation requires, which is a reading of the file rather than a test anyone ran. Nothing on this page was runtime tested against AutoGPT, and Kunavo publishes no AutoGPT integration guide.
Inside agent graphs, no custom-endpoint field is exposed to the person building the graph — on any plan, because it is the code rather than the plan that decides. blocks/llm.py is 1,651 lines with zero occurrences of base_url, api_base or openai_compatible, and its credential type is an api_key over a closed literal of eight named providers. The per-provider HTTP it delegates to lives in backend/util/llm/providers.py, and that is where to check rather than stopping at the block file: every base URL there is a hard-coded constant per provider — OpenRouter, the Llama API, AI/ML API, v0 — and the OpenAI client is constructed with no base URL argument at all. None of them is a value a graph author can set. The block-level ollama_host field is not a way around it either — it feeds Ollama's native client rather than a /v1 path, providers.py raises ValueError("Ollama does not support tools.") as soon as a call passes tools, and the Ollama page states the integration "is only available when self-hosting" and "cannot be used with the cloud-hosted version". Adding a model there means editing backend source: the Ollama page's own instructions are to add a catalog entry in data/llm_registry/catalog.py and a matching name line in data/llm_registry/llm_models.py, with an import-time check that "refuses to boot if an enum name has no catalog entry".
Kunavo serves no embedding model, so the embedding half of any of these tools stays with whoever you use for it today — that matters here because AutoGPT's marketplace search pins STORE_EMBEDDING_MODEL to 1536-dimension vectors, with the pgvector column declared vector(1536) in the schema and other dimensions failing on insert.
Maintained alternatives, with the licence and deployment boundary
| Project | Licence | Status on September 21, 2026 | Custom OpenAI-compatible endpoint |
|---|---|---|---|
| OpenHands | MIT | Active; v1.20.0 published September 17, 2026 | Settings, under Advanced: a Base URL field; the docs say it connects to "any LLM supported by LiteLLM" |
| Agent Zero | MIT (LICENSE file) | Active; pushed September 20, 2026 | An "Other OpenAI compatible" provider in conf/model_providers.yaml |
| CrewAI | MIT | Active; pushed September 21, 2026 | LLM(custom_openai=True, base_url=...), or the OPENAI_BASE_URL variable |
| Dify | GitHub reports NOASSERTION — read the licence file before forking | Active; pushed September 21, 2026 | An OpenAI-API-compatible provider plugin; field names not confirmed here |
| n8n | Sustainable Use License — internal business or non-commercial use only; not open source | Active; pushed September 21, 2026 | A Base URL field on the OpenAI credential, default https://api.openai.com/v1, confirmed in the credential source file rather than on the docs page |
| AutoGPT Platform, self-hosted | PolyForm Shield 1.0.0 | Active; v0.8.0 published September 19, 2026 | AutoPilot chat only — see the boundary section above |
Every endpoint claim in that table is configuration-surface reading — the vendor's own documentation, licence file or source file — and not a compatibility test of any kind.
Four names that recur on AutoGPT-alternatives lists should not be picked today. reworkd/AgentGPT is archived with its last push on April 29, 2025. FlowiseAI/Flowise announced end of life: feature development ceased July 29, 2026, the repository was archived August 13, 2026, and official team presence concluded August 31, 2026 — the Apache-2.0 code remains on GitHub to fork (see Flowise alternatives). microsoft/autogen states it "is now in maintenance mode" and sends new users to Microsoft Agent Framework. And SuperAGI is not archived but has published no release since v0.0.14 in January 2024 and no push since January 2025 — dates, not an announcement.
What migrates, what does not, and how to roll back
There is no documented AutoGPT Classic to AutoGPT Platform migration path. A recursive listing of all 8,730 paths on master on September 21, 2026 turned up no Classic-to-Platform converter and no import command — the paths matching "classic" outside classic/ are nine CI workflows, the docs/content/classic tree and one Platform admin UI component, none of them a conversion tool, and the migration scripts under the Platform backend are database migrations. The Classic README recommends the Platform, which is advice, not a tool. Read that as a rebuild, not a conversion, whichever destination you pick.
What actually carries is small and worth listing before you start: the text of your task and role description, the files in your workspace directory, and the value of your own model API key. What does not carry is the .env itself — SMART_LLM, FAST_LLM, RESTRICT_TO_WORKSPACE, the command-authorisation keys and OPENAI_API_BASE_URL have no one-to-one destination. Re-enter credentials in the new tool's own credential store rather than copying a plaintext key file between machines, and rotate any key that has been in a Classic checkout while you are there.
On the Platform side, agent portability is documented in one direction only: a tutorial covers Import from File, and the pricing table lists agent import and export on Pro, Max and Self-Host alike. The export format, whether credentials travel inside an exported file, and whether an agent exported from the cloud imports into a self-hosted instance are all undocumented — treat those as open questions rather than assumptions. The pricing table also advertises import from n8n, Make.com and Zapier, and that one does exist in the source rather than only in marketing: the library import dialog carries an external-workflow tab whose copy reads "Upload a workflow exported from n8n, Make.com, Zapier, or any other" alongside a "Workflow URL" field, and the hook behind that dialog hands the uploaded file to AutoPilot — it finishes by routing the browser to /copilot?source=import&autosubmit=true. So it is an AutoPilot-mediated import rather than a file-format converter, and how faithfully it reproduces a workflow is not something this page tested.
Rollback is the cheap part and the reason to sequence it this way. Keep the Classic checkout pinned at its current commit and keep the old .env until one real task has finished end to end on the new route; Classic is MIT, so keeping a fork indefinitely is permitted. Two further cautions if the destination is a self-hosted Platform: v0.8.0 is a breaking upgrade for installs that ran on the published default secrets — the release notes say ENCRYPTION_KEY, UNSUBSCRIBE_SECRET_KEY and BETTER_AUTH_SECRET no longer ship with values and the backend "does not start" on the old default — and without an embedding backend the marketplace's hybrid search degrades to lexical-only.
What it costs to stay, and what it costs to move
| Option | Published price | What it does not include |
|---|---|---|
| AutoGPT Pro (cloud) | $50.00 per month billed monthly; $42.50 per month billed annually | Agent runs, billed from a pre-paid credit wallet whose unit rate is not published |
| AutoGPT Max (cloud) | $320.00 per month billed monthly; $272.00 per month billed annually | The same wallet spend; the card reads "8.5x usage" for AutoPilot chat |
| AutoGPT Team (cloud) | "Coming soon — Contact us" | Not purchasable; no number published |
| AutoGPT Self-Host | No licence fee; "Open source", "BYO compute" | Your own model keys and infrastructure; support is "Community / GitHub" |
| OpenHands Open Source / Individual | Free / Free | Individual is 1 user and 10 maximum daily conversations; Enterprise is custom pricing |
| Dify Cloud Sandbox | Free | 200 message credits, 1 team member, 5 apps; paid tiers start above that |
AutoGPT prices were read from the pricing page's own data-monthly and data-yearly attributes on September 21, 2026, because the page renders the annual figure by default. Two official statements conflict on a free tier and you should know which one to trust: the pricing page's HTML meta description says "Start free, upgrade when your agents need more", while the homepage body says "The hosted cloud doesn't have a free tier, and that's deliberate". There is no free account card on the pricing page — only Pro, Max, Team and Self-Host, plus an instant demo with no signup. Confirm at checkout before you plan around either sentence. OpenHands and Dify figures come from openhands.dev/pricing and dify.ai/pricing, same date. n8n Cloud prices are localised by region and were captured only on the annual toggle in this research, so read them on n8n's own page rather than from any figure quoted second-hand.
The comparison you cannot complete is the honest finding here. AutoGPT's subscription gates AutoPilot chat volume, while running agents is pay-as-you-go from the credit wallet: the pricing page states outright that automations are "pay-as-you-go via the credit wallet — the same on every plan", and that "the subscription (Pro vs Max) gates AutoPilot Chat usage". What that page does not carry is a per-credit rate, a minimum top-up or an example run cost. So "subscription versus tokens" has no total you can compute from the published AutoGPT figures, and a per-run number quoted elsewhere is worth confirming with the vendor before you budget from it.
What can be modelled is the half you would move to a gateway: the self-hosted AutoPilot chat. Assume a working day of 40 turns at 25,000 uncached input tokens and 900 output tokens each — 1,000,000 input and 36,000 output in total. That ratio is an assumption for illustration, not a measurement of anyone's AutoPilot usage. Rates are live Kunavo catalog prices per million tokens.
| Model | Input / output per 1M | Estimated cost, one modelled day |
|---|---|---|
| Claude Haiku 4.5 | $0.40 / $2.00 | $0.472 |
| Gemini 3.8 Flash | $0.525 / $2.625 | $0.620 |
| GPT-5.6 Terra | $0.70 / $4.20 | $0.851 |
| Claude Sonnet 4.6 | $1.20 / $6.00 | $1.416 |
Across 22 such days that spread is $10.38 on Claude Haiku 4.5 against $31.15 on Claude Sonnet 4.6. This is illustrative token arithmetic, not a measured task cost and not a bill ceiling; it covers the chat turns only, and it does not include whatever your agent graphs spend through their own provider credentials. Kunavo's catalog amount is a billing floor rather than a cap: when the upstream reports its charge, the bill is the greater of catalog cost and upstream cost times the applicable markup. Cache charges and external tools sit outside this example. The minimum top-up is $10 in prepaid credit, a funding minimum rather than a task fee or a subscription — see billing details.
If the route you land on is OpenHands, Kunavo publishes a setup reference for it — a configuration guide, not a compatibility test. Start at the OpenHands integration, or read how the OpenAI-compatible endpoint works if you are wiring a client that has no page here, and create a Kunavo account when you are ready to fund a key. For the budgeting method rather than the rates, see AI cost optimization.
FAQ
Is AutoGPT Classic unsupported?
Yes, and the project says so in its own words. The README in the classic/ directory of the AutoGPT repository reads "This project is unsupported, and dependencies will not be updated" and "It was an experiment that has concluded its initial research phase." It adds that the codebase "has known vulnerabilities and issues with its dependencies" and should be used "for educational purposes only", and points readers to the AutoGPT Platform instead (checked September 21, 2026). The Classic setup documentation is still published and still resolves, which is why people keep starting an install that nobody maintains — a live docs page is not a statement of support.
Is AutoGPT dead or still maintained?
Half of it is finished and half of it is not, which is why a single yes or no is always wrong. AutoGPT Classic is unsupported by its maintainers' own README. The AutoGPT Platform, in the autogpt_platform/ directory of the same repository, published release autogpt-platform-beta-v0.8.0 on September 19, 2026, after v0.7.4 on September 4 and v0.7.3 on August 28 — six tagged releases since August 5. The GitHub API reported the repository as not archived and not disabled, with a push on September 21, 2026 (all checked September 21, 2026). Name the half you mean in every sentence, because the licence differs too: classic/ is MIT and autogpt_platform/ is PolyForm Shield 1.0.0.
What are the best maintained AutoGPT alternatives in 2026?
It depends on what you were using AutoGPT for, and the first filter is lifecycle rather than features. For an autonomous coding agent, OpenHands is MIT, published v1.20.0 on September 17, 2026, and exposes a Base URL field for any OpenAI-compatible endpoint. For a general-purpose agent with a Linux sandbox, Agent Zero is MIT and ships an "Other OpenAI compatible" provider. For agents written in Python, CrewAI is MIT and takes an explicit base_url. For visual workflow building, Dify and n8n are both active, but neither is plain open source — read each licence before you assume a fork is permitted. Three names that recur on AutoGPT-alternatives lists should not be picked today: reworkd/AgentGPT is archived, FlowiseAI/Flowise was archived on August 13, 2026, and microsoft/autogen states in its README that it is in maintenance mode with Microsoft Agent Framework as its successor (all checked September 21, 2026).
Can I move an AutoGPT Classic agent to the AutoGPT Platform?
Not by import. A search of the AutoGPT documentation and a scan of the repository tree on September 21, 2026 found no Classic-to-Platform converter, no migration command and no documented procedure — the Classic README only recommends the Platform, which is advice rather than a tool. A Classic agent is rebuilt by hand as a Platform block graph, or rebuilt in whatever tool you move to instead. What actually carries over is the text of your task and role description, the files in your workspace directory, and the value of your own model API key; the .env variables themselves do not map, because the Platform has no block-layer equivalent for OPENAI_API_BASE_URL. Plan the rebuild as work, not as a conversion step.
Can I point AutoGPT at my own OpenAI-compatible endpoint?
Only in one place, and only on a self-hosted install. AutoGPT's own documentation for the AutoPilot chat says it can run "without an Anthropic, OpenAI, or OpenRouter key" by routing through "any OpenAI-compatible HTTP endpoint you control", and its scenario table explicitly includes "A managed OpenAI-compatible API you don't pay AutoGPT for" pointed at its /v1 URL. That is configured with CHAT_USE_LOCAL, CHAT_BASE_URL and CHAT_API_KEY, and the same page states that the cloud deployment at agpt.co ignores those variables entirely. The LLM blocks inside agent graphs are a different story on every plan: backend/blocks/llm.py contains no base_url or api_base field at all, and its credential type is an api_key over a closed list of eight named providers (checked September 21, 2026).
How much does AutoGPT cost compared with self-hosting?
On agpt.co the pricing page showed Pro at $50.00 per month billed monthly or $42.50 per month billed annually, and Max at $320.00 monthly or $272.00 annually, with Team listed as "Coming soon — Contact us" (read from the page's own data-monthly and data-yearly attributes on September 21, 2026). Self-hosting carries no licence fee: the Self-Host column reads "Open source" with "BYO compute", community support through GitHub, and a note that all model providers require your own API keys. The figure nobody can budget from is the agent-run side — AutoGPT bills automations from a pre-paid credit wallet, and the pricing page that describes that wallet gives no per-credit rate, no minimum top-up and no example run cost. Treat any per-run number you find for AutoGPT as something to confirm with the vendor rather than as a published figure.
AutoGPT repository files, release metadata, documentation and the agpt.co pricing and homepage were read on September 21, 2026, as were the GitHub API records and pricing pages for OpenHands, Agent Zero, CrewAI, Dify, n8n, AgentGPT, Flowise, AutoGen and SuperAGI. Every compatibility statement on this page is source-document and configuration reading; no client named here has been runtime tested against Kunavo. Kunavo token rates come from the live catalog, and every dollar example is illustrative token arithmetic.