What is CoreLayer?

CoreLayer is a local-first desktop Agent OS (AI Agent Operating System) designed to manage Agents, local tools, execution states, and personal workspaces.

The built-in AI persona is named Jarvis. It is not another chatbot skin, nor is it a simple IDE plugin. It establishes a secure control plane between your personal applications, local operating system, LLMs, and MCP servers using strict permission rules, isolated sandbox sessions, and verifiable execution logs.

Product Vision

1. From "Chat History" to "Execution System"

Most existing AI tools bury tasks inside a single chat thread. As context grows, managing state becomes impossible. CoreLayer structures tasks into a manageable execution system:

• You prompt Jarvis with a high-level goal.
• CoreLayer resolves this goal into a Workspace, which manages a Task Graph (task dependencies), Agent Runs (execution sessions), scoped approvals, audit trails, and final deliverables (Artifacts).

Goal
  ↓
Workspace
  ├─ Task Graph
  ├─ Agent Runs
  ├─ Scoped Approvals
  └─ Sandbox Sessions (Environment)
  ↓
Verifiable Deliverables (Artifacts) & Execution Trajectories

2. Executor Governance (Claude Code / Codex / OpenCode)

While AI coding agents like Claude Code, Codex, and OpenCode are highly capable, executing arbitrary shell commands directly on your host machine poses severe security risks, and their lifecycle remains untracked. CoreLayer provides adapters to run them as Managed Executors:

• Sandboxed Sessions: Changes are isolated inside a git-worktree or containerized sandbox prior to committing.
• Granular Approvals: High-risk operations (such as file deletes, writes, or outbound network calls) are intercepted and suspended by the Permission Guard, prompting manual user verification.
• Deliverables Separation: Final deliverables (Artifacts like code diffs, specs, or reports) are clean-filtered and separated from compile noise, test logs, and terminal streams.

3. "Jarvis Manages the Work; Specialized Tools Execute it"

This is the core design philosophy of CoreLayer. We do not reinvent the wheel—we provide the orchestrator. Jarvis coordinates tasks and enforces security rules, while specialized tools or MCP services execute the operations. This ensures that the system is not locked to any single model provider or development environment.

Core Pillars

• MCP-First Integration — Natively supports stdio, HTTP, and SSE transport protocols, registering external tools under the unified mcp:{serverId}:{toolName} namespace.
• Permission Guard — Classifies tool invocations into Low, Medium, High, and Critical risk levels, using decision caches (Permission Memories) to prevent alert fatigue, and records a complete audit log.
• Unified Tool Registry — Collects and manages native Rust modules, MCP tools, local JS/TS skills, and external REST endpoints.
• Model Gateway & Routing — Dynamically routes LLM queries across MiMo, Groq, DeepSeek, OpenAI, Gemini, Qwen, and Ollama, matching endpoints to latency, tool-calling, or reasoning requirements.
• Voice-Native Pipeline — Features wake-word detection, ASR transcription, streaming sentence-level TTS speech generation, and barge-in audio interruption.
• Local-First Tri-Mode Storage — Uses a zero-configuration SQLite database locally by default, with settings supporting cloud sync using Supabase or self-hosted PostgreSQL.

Next Steps

• Install CoreLayer — Download and run the desktop client.
• Quickstart Guide — Connect your first model and MCP server.
• System Architecture — Understand how the Tauri Shell and Node.js daemon cooperate.