Only 50 of 272 AI Agent Projects Can Be Self-Hosted
This is a count, not an opinion. Every number below comes from the hedirbase.com index of 1,444 approved entries, filtered to the three AI agent categories and then checked entry by entry for a working GitHub repository. Nothing here is estimated.
"AI agent" is one label covering three different business models, and the difference decides whether you can run the thing yourself. Frameworks are open because adoption is the product. Products are closed because access is the product. Platforms are ambiguous because the word gets used for both. In our index the split is 40%, 14%, and 7% — so any list that ranks all three together is comparing three kinds of animal.
How the count was made
Three steps, all reproducible:
- Filter the index to the three agent categories:
AI Agent Frameworks(75),AI Agent Platforms(85),AI Agent Products(112). Total 272. - Count entries with a non-empty
repo_urlfield. That returned 54. - Inspect all 54 by hand and remove entries whose repo link does not actually point at independent deployable code.
Step 3 removed four entries, and the reasons matter for anyone reusing this data:
| Entry | Why removed |
|---|---|
| AutoGen | Indexed twice, same repo (microsoft/autogen) |
| Pydantic AI | Indexed twice, same repo (pydantic/pydantic-ai) |
| Magentic-One | Repo field points at microsoft/autogen; it is a component of AutoGen, not a separate project |
| Copilot Workspace | Repo field points at github.com/features/copilot, a product page, not a repository |
54 minus 4 equals 50. If you see 54 quoted elsewhere, including in our own raw export, that number double-counts two projects and includes one pointer to a marketing page.
The three categories are not equally open
| Category | Indexed | Self-hostable | Share |
|---|---|---|---|
| AI Agent Frameworks | 75 | 30 | 40.0% |
| AI Agent Platforms | 85 | 12 | 14.1% |
| AI Agent Products | 112 | 8 | 7.1% |
| Total | 272 | 50 | 18.4% |
The pattern is consistent with how these things are built. A framework is a library: its value comes from being adopted by developers, so the code is public by default. A product is a business: its value comes from being the only place the workflow runs, so the code stays closed. Platforms sit in between because "platform" is used for both.
If you are evaluating an "AI agent platform", the category label is a weak signal and the repo field is a strong one. Two products with similar marketing pages can differ completely on whether you can run them on your own hardware.
The 50 that can be self-hosted
Frameworks — 30 entries
| # | Project | Repository |
|---|---|---|
| 1 | Agent Zero | agent0ai/agent-zero |
| 2 | AgentForge | DataBassGit/AgentForge |
| 3 | Agno | agno-agi/agno |
| 4 | Apache Burr | apache/burr |
| 5 | Atomic Agents | Eigenwise/atomic-agents |
| 6 | AutoGen | microsoft/autogen |
| 7 | BabyAGI | yoheinakajima/babyagi |
| 8 | Browser Use | browser-use/browser-use |
| 9 | ChatDev | OpenBMB/ChatDev |
| 10 | Claude Agent SDK | anthropics/claude-agent-sdk |
| 11 | CopilotKit | CopilotKit/CopilotKit |
| 12 | DSPy | stanfordnlp/dspy |
| 13 | Daytona | daytonaio/daytona |
| 14 | Dify | langgenius/dify |
| 15 | Eidolon | eidolon-ai/eidolon |
| 16 | Eliza | elizaOS/eliza |
| 17 | Guardrails AI | guardrails-ai/guardrails |
| 18 | Guidance | guidance-ai/guidance |
| 19 | Julep | julep-ai/julep |
| 20 | Langflow | langflow-ai/langflow |
| 21 | Letta | letta-ai/letta |
| 22 | Marvin | PrefectHQ/marvin |
| 23 | MetaGPT | FoundationAgents/MetaGPT |
| 24 | OpenAI Agents SDK | openai/openai-agents-python |
| 25 | Phidata | phidatahq/phidata |
| 26 | Pydantic AI | pydantic/pydantic-ai |
| 27 | Semantic Kernel | microsoft/semantic-kernel |
| 28 | Smolagents | huggingface/smolagents |
| 29 | Stagehand | browserbase/stagehand |
| 30 | Strands Agents | strands-agents/sdk-python |
Platforms — 12 entries
| # | Project | Repository |
|---|---|---|
| 1 | Agenta | Agenta-AI/agenta |
| 2 | Agentset | agentset-ai/agentset |
| 3 | AutoGPT | Significant-Gravitas/AutoGPT |
| 4 | Automatisch | automatisch/automatisch |
| 5 | Cognee | topoteretes/cognee |
| 6 | Hexabot | hexabot-ai/hexabot |
| 7 | Onyx | onyx-dot-app/onyx |
| 8 | Rivet | Ironclad/rivet |
| 9 | Sim Studio | simstudioai/sim |
| 10 | SuperAGI | TransformerOptimus/SuperAGI |
| 11 | Superagent | homanp/superagent |
| 12 | Wren AI | Canner/WrenAI |
Products — 8 entries
| # | Project | Repository |
|---|---|---|
| 1 | Cline | cline/cline |
| 2 | Codex CLI | openai/codex |
| 3 | Devika | stitionai/devika |
| 4 | Kortix Suna | kortix-ai/suna |
| 5 | Nanobrowser | nanobrowser/nanobrowser |
| 6 | PyGPT | szczyglis-dev/py-gpt |
| 7 | Vane (formerly Perplexica) | ItzCrazyKns/Vane |
| 8 | Black Forest Labs FLUX | black-forest-labs/flux |
What this count does not tell you
Three limits, stated plainly:
- Open repo does not mean free to run. Several of these projects are open-core: the repository is public, but production features sit behind a paid tier. Hosting also costs compute, which for agent workloads is not trivial.
- "Can be self-hosted" is judged by the presence of a repository, not by whether we deployed it. A public repo with a broken install guide would still be counted here. We did not attempt a deployment for each of the 50.
- The index is not the market. These 272 entries are what has been submitted and approved to hedirbase.com, which skews toward projects with an English web presence. The true ratio across all agent projects is unknown and probably lower, because closed products rarely submit themselves to directories.
Use the field, not the label
If you take one thing from this: when a list calls something an "AI agent platform", that word has told you almost nothing about whether you can run it yourself. The presence of a repository is the fact. Everything else is positioning.
The 50 projects above are all reachable from the hedirbase.com index, with descriptions and category placement. If a project you rely on is missing from the list, it is either closed source or not yet in the index — you can check which by searching the site.