I'm a student from Bangladesh. While the biggest companies in the world race to build AI agents, I kept asking myself one question:
Can one person, with one laptop and no team, build a real AI agent engine?
I tried. The answer is ZYRAXON AI: an open-source desktop agent that doesn't just answer. It does the work.
A chatbot gives you words. ZYRAXON takes a goal and carries it all the way to a result.
flowchart LR
A([🎯 Your goal]) --> B[🧭 Plan]
B --> C[🧰 Pick tools]
C --> D[⚙️ Execute]
D --> E[👁️ Observe]
E --> F{✅ Done?}
F -- not yet --> B
F -- yes --> G([🏁 Result])
It plans the steps, picks the right tools, runs them, reads what came back, fixes its own mistakes, and checks the job is actually finished before it tells you it is.
| 🧰 | 🔌 | 🎛️ | 🧠 |
|---|---|---|---|
| 888 tools | 8,202 MCP servers | 11 agent modes | Your model |
| across 63 categories | one catalog, one place | plan · build · explore · auto and more | plug in the brain you prefer |
| 🖥️ | 🤝 | 🌍 | 🔓 |
|---|---|---|---|
| Desktop · Terminal · SDK | Sub-agents | Control anything with an interface | Open source |
| three ways to use it | parallel agents under one supervisor | commands, APIs, tools | read it, fork it, improve it |
If something can be controlled through commands, APIs or tools, a ZYRAXON agent can drive it. The few things that truly need human hands stay with a human, and the agent handles everything else.
MCP servers are the agent's hands. More hands, more power. The ZYRAXON MCP Hub puts 8,202 MCP servers in one catalog:
- 8,094 servers collected from the public MCP registry, each one tagged with how far it was tested
- 108 hand-curated apps (Notion, Linear, Atlassian and more) with setup steps and guides
- 17 categories and 64,000+ tools observed across the catalog
| 8,202 | 8,094 | 108 | 17 | 64K+ |
|---|---|---|---|---|
| total entries | from the public registry | hand-curated | categories | tools observed |
The full list lives in one file, generated by script from the catalog:
MCP Hub/ALL MCP.md
ZYRAXON shows you only the step that matters:
| If the server needs… | Servers | What you'll see |
|---|---|---|
| ⚡ Nothing at all | 4,679 | Press Connect. The tools are live. |
| 🌐 A browser sign-in | 3,105 | A sign-in page opens, you approve, you're in. |
| 🔑 An API key | 228 | A short form asks for the key. |
| 🔗 Your own address | 82 | You paste your instance URL and go. |
%%{init: {"theme": "base", "themeVariables": {"pie1": "#0891B2", "pie2": "#8B5CF6", "pie3": "#38BDF8", "pie4": "#312E81", "pieTitleTextSize": "18px", "pieSectionTextColor": "#ffffff"}}}%%
pie showData title How the 8,094 registry servers connect
"No sign-in needed" : 4679
"Browser sign-in" : 3105
"API key" : 228
"Your own address" : 82
Nothing here is marked "works" on faith. Every registry server carries a label for how far the check went:
| Proof level | Servers | What was proven |
|---|---|---|
| ✅ full | 2,989 | initialize, tools/list and a read-only tools/call all answered |
| 🟢 tools_list | 1,391 | initialize and tools/list answered |
| 🔐 auth_discovery | 3,105 | a 401 was followed by a working sign-in discovery document |
| 🚀 launched | 228 | the package resolved, started, then asked for its key |
| 📍 declared | 82 | the endpoint is reachable but needs a value only you have |
| ⏳ pending | 299 | local packages listed, proof not yet recorded |
7,795 of the 8,094 registry servers already carry recorded proof. The remaining 299 are being worked through, and the label on each row tells you exactly where it stands.
%%{init: {"theme": "base", "themeVariables": {"pie1": "#0891B2", "pie2": "#8B5CF6", "pie3": "#38BDF8", "pie4": "#312E81", "pieSectionTextColor": "#ffffff"}}}%%
pie showData title Install kinds
"Remote HTTP (nothing to install locally)" : 7567
"npx (npm package)" : 217
"uvx (PyPI package)" : 185
"docker (container image)" : 125
📂 See all 17 categories
| Category | Servers |
|---|---|
| AI & ML | 2,293 |
| General | 2,076 |
| Web & Search | 1,367 |
| Developer Tools | 731 |
| Communication | 327 |
| Data & Analytics | 277 |
| Finance & Crypto | 213 |
| Productivity & Docs | 164 |
| Design & Media | 149 |
| Maps & Travel | 121 |
| Databases | 95 |
| Business & CRM | 79 |
| Cloud & Infrastructure | 64 |
| Security | 51 |
| Knowledge & Memory | 44 |
| Files & Storage | 34 |
| IoT & Hardware | 9 |
Important
The ZYRAXON Model Provider is paused on purpose. We will switch on our own provider only when this repository reaches 100,000 GitHub stars. Until then, the provider endpoint replies with a short notice saying exactly this.
Want it sooner? Star the repo and share it. Every star moves the date closer.
The ten-model lineup (Apex · Matrix · Pulse · Titan · Neural · Codex · Visionary · Cyber · Quantum · Genesis) is waiting for that day.
Meanwhile you lose nothing: the ZYRAXON agent works with the model you bring.
I don't want to carry this alone. ZYRAXON is open source, and the next chapter is better when we write it together.
| 🧪 Test | Connect a few MCP servers and tell me what breaks |
| 🐛 Report | Open an issue for any bug or idea |
| 🔌 Build | Send a pull request with a tool or an integration |
| 📝 Write | Improve the docs, in English or Bengali |
| ⭐ Spread | Leave a star so more people can find it |
Student, senior engineer, or just curious: there is a place for you here.
GitHub · Website · MCP Hub · Agent Ecosystem Hub
ZYRAXON AI © 2026 · See the repository for license and terms.