Note
Pre-Release Software - AgentSystems is in active development. Join our Discord for updates and early access.
Self-hosted app store and runtime for third-party AI agents.
Most AI companies are building for the wrong future. AI distribution is inverting: instead of sending data to agents, agents will run where data lives.
Why the inversion is inevitable:
- Frontier models, open-source models, and local hardware all keep improving
- Building agents is quickly becoming trivial
- Models can run locally or via your own cloud accounts — reducing dependency on third-party agent platforms
- NVIDIA's research shows small language models excel at specialized tasks — most agent workloads don't need massive generalist models
- As AI transforms white-collar work, millions of subject matter experts (accountants, analysts, consultants, underwriters) will create agents to leverage their domain expertise. A tax accountant builds an agent for startup equity compensation. An insurance underwriter builds one for risk assessment.
- The market floods with specialized agents for every niche task
Two problems emerge: discovery and trust. When agents are everywhere, how do you discover them? How do you run third-party agents without exposing your data or credentials?
AgentSystems provides infrastructure for both:
Distribution:
- Self-hosted catalog indexed by GitHub username
- Federated Git-based discovery with auto-merge
- Permissionless publishing
Security model:
- Container isolation
- Runtime credential injection (separated from agent provider)
- Default-deny egress with per-agent allowlists
- Provider abstraction (agents inherit your OpenAI/Anthropic/Bedrock/Ollama config)
- Hash-chained audit logs
Infrastructure for a commoditized agent future.
📺 Watch the 100-second demo | Full walkthrough (9 min)
# macOS / Linux
curl -fsSL https://github.com/agentsystems/agentsystems/releases/latest/download/install.sh | sh
agentsystems init agent-platform && cd agent-platform
agentsystems upThat's it. Open http://localhost:3001 and start discovering agents.
Prefer manual installation? See the manual installation guide.
What's running: Gateway (API), UI (web interface), egress proxy (network controls), and PostgreSQL (audit logs).
Next: In the UI, click "Discover" → Find an agent → Click "Add" → Invoke it
Have an NVIDIA GPU? See the GPU acceleration guide.
1. Build agent with model requirements |
2. Push to container registry |
3. List your agent in the index (automated merge) |
4. Agent available for discovery |
graph TB
subgraph "Your Infrastructure"
App[Your Application]
UI[Web UI :3001]
GW[Gateway :18080]
Agent[Third-Party Agent<br/>Container]
Proxy[Egress Proxy :3128]
Storage[Thread Storage<br/>/artifacts/thread-id/]
Audit[Hash-Chained Logs<br/>PostgreSQL]
end
subgraph "External"
Index[Agent Index<br/>agentsystems.github.io]
Allow[approved.com ✓]
Block[evil.com ✗]
AI[AI Providers<br/>OpenAI/Anthropic/Ollama]
end
Index -.->|1. Discover & Pull| GW
UI -->|Manage| GW
App -->|2. POST /invoke/agent| GW
GW -->|3. Route + Inject X-Thread-Id| Agent
Agent -->|4. Outbound Call| Proxy
Proxy -->|Check Allowlist| Allow
Proxy -.->|Blocked| Block
Agent <-->|5. Read/Write| Storage
Agent -->|6. Call LLM| AI
Agent -->|7. Return Result| GW
GW -->|8. Response| App
GW -->|Audit Logs| Audit
style Agent fill:#1a4d5c
style Proxy fill:#2d5f3f
style Block fill:#5c1a1a
style Audit fill:#3d3d5c
style UI fill:#3d5c4d
Agent frameworks (LangChain, Agno, CrewAI)? Those help you build agents. AgentSystems helps you discover and run agents built by others. You can build with those frameworks and publish to AgentSystems.
Cloud-hosted agents (Harvey, Artisan)? Your data goes to their servers. AgentSystems runs on your infrastructure — designed to keep your data under your control.
Deployment platforms (Coolify, Heroku)? Those deploy web apps. AgentSystems deploys AI agents with agent-specific features: model routing, egress controls, thread-scoped storage, federated discovery.
Manual Docker / Docker Compose? AgentSystems adds discovery and pre-built infrastructure: federated agent catalog, UI, gateway, egress proxy, audit logs.
The platform consists of 6 interdependent repositories:
| Component | What It Does | Technology | Latest Version |
|---|---|---|---|
| agent-control-plane | Gateway, orchestration, egress proxy | FastAPI, PostgreSQL | |
| agentsystems-sdk | CLI tool for deployment | Python, Docker | |
| agentsystems-ui | Web interface for management | React, TypeScript | |
| agentsystems-toolkit | Library for building agents | Python, LangChain | |
| agent-template | Starter template for new agents | FastAPI, LangGraph | rolling |
| agent-index | Federated agent discovery | GitHub Pages | rolling |
Q: How does agent publishing work?
A: Fork the agent-index repo, create developers/your-github-username/, add your agent metadata, and submit a PR. Auto-merge validates that the folder name matches your GitHub username (fork ownership = namespace ownership). No manual approval needed.
Q: Are agents reviewed before being listed?
A: No. The AgentSystems Community Index lists agents from third-party developers. AgentSystems does not review or endorse software in any index.
Q: Can I use this in production?
A: AgentSystems is pre-release. Use for development and testing. For production, wait for stable release or contact us on Discord for early access guidance.
Q: Can I run a private agent index?
A: Yes. The index is Git-based and federated. Fork the agent-index repo to run your own index.
Getting Started:
For Agent Developers:
We welcome contributions across the stack:
- 🤖 Build Agents - Create specialized agents and publish to the index
- 🔒 Security - Improve isolation, audit mechanisms, or egress controls
- 📚 Documentation - Write guides, tutorials, or API references
- 🐛 Bug Reports - Help identify and fix issues
See CONTRIBUTING.md for guidelines.
- Discord - Chat with developers and contributors
- GitHub Issues - Bug reports and feature requests
Licensed under the Apache-2.0 license.