Self-hosted app store for AI agents. Federated discovery, container isolation, run on your infrastructure.
The problem: most organizations either build every agent in-house or send their data to third-party servers.
You wouldn't build your own email client or maps app - you'd download one. AI agents should work the same way. I spent a year building the infrastructure to make that possible: discover agents built by others, run them on your infrastructure (private cloud, on-premises, or local) without sending your data to third-party servers.
Key architecture:
Federated Git-based index (fork-based ownership, no gatekeepers)
Container isolation + egress proxy (you configure which URLs agents can access)
Credential injection (API keys configured on host, not in agent images)
Model abstraction (works with Ollama local, cloud APIs, or hybrid)
Hash-chained audit logs
The platform works. The agent index is mostly empty, but someone has to build the rails so the trains can run.
Apache-2.0 open source. Pre-release but functional.
GitHub: https://github.com/agentsystems/agentsystems
Docs: https://docs.agentsystems.ai
Looking for:
Agent builders to publish to the index
Security researchers to review the architecture
Organizations that need self-hosted AI infrastructure
Top comments (1)
Aloha! Author here.
Origin story:
I built this after repeatedly seeing the same problem: companies wanted to use specialized AI agents (document processing, compliance checks, data analysis) but couldn't justify either: (1) building everything in-house - expensive and creates unmaintainable internal tools, or (2) sending sensitive data to third-party SaaS platforms.
The same pattern kept repeating: they needed "an app store for agents" that worked like Docker Hub but with security isolation baked in.
The interesting technical choice:
I wanted this to be fully open with no gatekeepers - no accounts, nothing to signup for, and as decentralized as possible. So I came up with a federated Git-based index. Instead of a centralized marketplace, agent metadata lives in a GitHub repo. Developers fork it, add their agent to
developers/their-username/, and submit a PR. Automated validation checks that the folder name matches the fork owner (GitHub-authenticated namespace ownership). No manual approval needed. The entire system operates without gatekeepers - fully forkable if you disagree with the index maintainers.Current state & what I need:
Pre-release and functional. The platform works, but the agent index is mostly empty since we're just getting started. I'm looking for:
Agent builders - If you've built agents for internal use, consider publishing them. The SDK makes it straightforward.
Security researchers - I want this architecture stress-tested. Found a hole? I'll credit you and fix it.
Organizations with data governance requirements - If you need self-hosted AI infrastructure, let's talk about your use case.
Tech stack:
Control plane: FastAPI, PostgreSQL, Docker SDK
UI: React, TypeScript, Zustand
SDK: Python, Typer
Agents: Any language (FastAPI template provided)
Happy to answer questions about the architecture, federated index mechanics, or security model. Critiques are welcome too.