The open-source
agent operating system

Claude is the kernel. Every AI tool is a process. All agents share one memory.
One daemon routes your tasks across Claude, local Ollama models, and the CLI agents you already use — with plug-in MCP drivers and semantic memory that crosses agent boundaries.

one daemon·many agents·shared memory·MCP drivers
# Claude Code writes some code… $ claudeos run "write a one-line python script to reverse a string" [claude-code] print(input()[::-1]) # …and a fully local model remembers it. Different backend, same memory. $ claudeos run "what script did the coding agent write earlier?" --provider ollama [ollama / llama3.2] print(input()[::-1])

An OS, not another chatbot

ClaudeOS treats your AI tools the way an operating system treats processes.

🧬

Kernel & scheduler

A daemon routes every task to the right backend via configurable rules — coding to Claude Code, quick summaries to a local model, deep reasoning to the Claude API.

⚙️

Agents as processes

Claude Code, Gemini CLI, Codex, and any Ollama model wrapped as first-class providers. Nothing reimplemented — orchestrate the tools you already use.

🔌

MCP drivers

Capabilities plug in as MCP servers. claudeos mcp install fetch and the kernel model can browse the web; filesystem, GitHub, Puppeteer and more in the registry.

🧠

Shared semantic memory

Every run is embedded locally and injected into future tasks — whichever agent produced it. SQLite + Obsidian-compatible markdown, fully private.

🔗

Cross-provider pipelines

research → implement → verify chains where each step runs on a different backend, streamed step-by-step over WebSocket.

🖥

Native desktop app

A 2.8 MB Tauri dashboard: run console, live event feed, memory browser, config editor. The CLI and the app are both thin clients over one local API.

The OS analogy, taken seriously

 

OS conceptClaudeOS
Kernel / schedulerDaemon that routes tasks to backends via config rules
ProcessesExisting CLI agents wrapped as providers
DriversMCP servers, installed from a curated registry
Shared memorySQLite index + markdown vault every agent reads & writes
System callsLocal HTTP + WebSocket API any client can attach to

Install

Node ≥ 23.6 runs the TypeScript natively — no build step.

# the daemon
git clone https://github.com/RishavRaj20/claudeos && cd claudeos
npm install
cp claudeos.config.example.json claudeos.config.json
node src/cli.ts start

# then, in another terminal
node src/cli.ts run "explain CRDTs in 3 lines"
node src/cli.ts pipeline build "a python function is_palindrome(s)"

Optional: Ollama for local models & semantic memory · ANTHROPIC_API_KEY for the direct API · the desktop app for a native dashboard.