┌──────────────────────────────────────────────────────────────────┐
│ │
│ You write a markdown file. │
│ The kernel compiles it into an agent. │
│ The agent uses tools, creates applets, learns skills. │
│ Everything runs in the browser — no backend. │
│ │
└──────────────────────────────────────────────────────────────────┘
LLMos is an operating system for AI agents. It treats language models the way Unix treats C programs — as the native executable format. Agents are markdown files. The kernel is markdown. Skills, memory, and configuration are all markdown.
LLMos is the prefrontal cortex of a three-part cognitive ecosystem:
| Repository | Brain Region | Role |
|---|---|---|
| evolving-memory | Hippocampus | Cognitive Trajectory Engine — dream consolidation, topological memory, ISA/VM |
| RoClaw | Cerebellum | Physical embodiment — vision loop, motor ISA, semantic navigation |
| llmos (this repo) | Prefrontal Cortex | Agent kernel — orchestration, applets, skills, workspace UI |
Book · Roadmap · Contributing
graph TB
subgraph KERNEL["LLMos Kernel"]
direction TB
AGENTS["Agent Manager<br/>Load, compile, orchestrate"]
SKILLS["Skills Manager<br/>Discover, filter, promote"]
APPLETS["Applet Runtime<br/>Generate, execute, display"]
STORAGE["Storage Layer<br/>IndexedDB filesystem"]
end
subgraph UI["Workspace UI"]
direction TB
CHAT["Chat Panel"]
SIDEBAR["Sidebar + Volumes"]
CONTEXT["Context Panel + Views"]
CMD["Command Palette"]
end
subgraph EXTERNAL["Sibling Systems"]
direction TB
MEMORY["evolving-memory<br/>Dream engine, traces, graph"]
ROCLAW["RoClaw<br/>Vision, motors, navigation"]
end
LLM["LLM Provider<br/>(Claude, OpenRouter)"]
UI --> KERNEL
KERNEL --> LLM
KERNEL -.->|"traces, dreams"| MEMORY
KERNEL -.->|"agent definitions"| ROCLAW
style KERNEL fill:#5b21b6,color:#fff
style UI fill:#1e3a5f,color:#fff
style EXTERNAL fill:#065f46,color:#fff
style LLM fill:#b45309,color:#fff
The kernel runs entirely in the browser. Agents are markdown files with YAML frontmatter — the development LLM creates, reads, modifies, and orchestrates them. When the system learns a new pattern, it writes it into a skill file. The agent definition is the documentation is the evolution history.
git clone https://github.com/EvolvingAgentsLabs/llmos
cd llmos
npm install
npm run devOpen http://localhost:3000. The workspace launches with a chat panel, sidebar, and context viewer — everything runs client-side.
For the Electron desktop app:
npm run electron:devTraditional OS LLMos
───────────── ─────
Compiled binaries → Markdown files
Processes → Agent execution loops
File system → Volume system (System / Team / User)
Permissions → Volume access + kernel rules
System calls → LLM inference requests
IPC → Agent-to-agent messaging
Package manager → Skill promotion pipeline
Skills flow upward as they prove reliable:
User Volume ──→ Team Volume ──→ System Volume
(personal workspace) (shared intelligence) (immutable foundation)
Promote at Promote at
5+ uses, 80% success 10+ uses, 90% success
The ClientKernel orchestrates all components:
import { getKernel } from '@/lib/kernel';
const kernel = getKernel(llmCallback);
// Create a project
const project = await kernel.createProject('my-project', 'Build a data analyzer');
// Ensure agents exist
await kernel.ensureProjectAgents(project.path, project.goal);
// Invoke an agent
const result = await kernel.invokeAgent(
'projects/my-project/agents/AnalyzerAgent.md',
'Analyze this dataset',
project.path
);Agents can generate runtime applets — small interactive components that render in the workspace. The LLM writes the code, the applet runtime executes it, and the desktop manager displays it.
llmos/
├── app/ # Next.js pages + layout
├── components/
│ ├── chat/ # Chat panel, messages, input
│ ├── workspace/ # Adaptive layout, panels, command palette
│ ├── sidebar/ # Sidebar, volume browser
│ ├── applets/ # Applet rendering
│ ├── artifacts/ # Artifact viewer
│ ├── layout/ # Header, navigation
│ └── system/ # 3D background, holographic UI
├── contexts/ # React contexts (Workspace, Project, Applet)
├── hooks/ # Custom hooks (code execution, workflows)
├── lib/
│ ├── kernel/ # ClientKernel — boot, orchestration, API
│ ├── agents/ # Agent compiler, loader, messenger, orchestrator
│ ├── skills/ # Skill manager, promotion pipeline
│ ├── applets/ # Applet store, desktop manager, error fixer
│ ├── llm/ # LLM client, types, storage
│ ├── storage/ # IndexedDB filesystem adapter
│ ├── debug/ # Logger, console store, metrics
│ ├── runtime/ # Applet runtime, speculative execution
│ └── ... # Tool executor, MCP client, utilities
├── public/
│ ├── system/ # Kernel rules, system agent definitions
│ └── volumes/ # System volume (skills, tools)
├── electron/ # Electron shell (serial, native features)
└── docs/ # The Book
| Layer | Technology |
|---|---|
| Development LLM | Claude Opus 4.6 via Claude Code |
| Frontend | Next.js 14, React, Tailwind CSS |
| Desktop | Electron (optional) |
| Kernel | TypeScript, client-side only |
| Storage | IndexedDB (via filesystem adapter) |
| Agent Format | Markdown + YAML frontmatter |
| Memory | evolving-memory (REST API) |
| Hardware | RoClaw (separate repo) |
LLMos is one part of a three-repository ecosystem. Each repo maps to a brain region with a distinct responsibility:
evolving-memory (Hippocampus) handles persistence and learning — the Cognitive Trajectory Engine with dream consolidation (SWS/REM/Consolidation phases), topological memory graphs, fidelity-weighted trace merging, and a REST/WebSocket server that any client can connect to.
RoClaw (Cerebellum) handles physical embodiment — the vision-language loop, hex-bytecode motor ISA, semantic navigation maps, and ESP32 firmware. It connects to evolving-memory via HTTP for dream cycles and trace ingestion.
llmos (Prefrontal Cortex) handles high-level orchestration — agent compilation, skill management, applet generation, and the workspace UI. It's the "thinking" layer that decides what to do, while RoClaw decides how to move and evolving-memory decides what to remember.
| Phase | Status |
|---|---|
| Phase 0: Distributed Instruction Runtime | Done |
| Phase 1: Desktop app, agent/volume/kernel system | Done |
| Ecosystem restructuring (Cognitive Trinity) | Done |
| evolving-memory REST server integration | Done |
| RoClaw memory deduplication | Done |
| Phase 3: Multi-agent workflows, advanced orchestration | In Progress |
| Phase 4: Plugin architecture, community skills | Planned |
See ROADMAP.md for detailed milestones.
The accompanying book walks through every layer of the system — from the philosophical thesis ("LLM as kernel") down to the TypeScript implementation.
Read it at evolvingagentslabs.github.io/llmos
LLMos is a research system. Expect architectural changes and experimental modules.
Current priorities:
- Multi-agent orchestration patterns
- Applet generation and runtime improvements
- Skill promotion pipeline refinement
- evolving-memory integration (traces, dream cycles)
- Agent and skill pattern contributions (markdown)
See CONTRIBUTING.md for guidelines.
Apache 2.0 — Built by Evolving Agents Labs.
A markdown file becomes an agent. The agent thinks with a language model. It creates tools, learns skills, and remembers through dreams. Three repos, one cognitive architecture. This is LLMos.