CoreLayerCoreLayer
Local-firstPermission-guardedMCP connected

CoreLayer

A local-first AI control layer for your desktop apps, tools, models, and MCP workflows.

Jarvis coordinates your personal apps through a permission-guarded desktop command center. Connect MCP servers, route work across models, approve risky actions, and use voice without giving up local control.

DownloadRead the docsView on GitHub
CoreLayer Desktop
CoreLayer desktop command center showing Jarvis, MCP tools, and system status
Local-first SQLite
MCP-first integrations
Permission-guarded tools
Voice-native workflows
Model routing
Tauri desktop app

Your apps are separate. Your assistant should understand the whole system.

Most assistants are trapped in chat windows, single-app copilots, or cloud automation platforms. CoreLayer sits locally between your personal apps, tools, models, and MCP servers so Jarvis can coordinate work through explicit permissions.

Without CoreLayer

Apps work in isolation
No unified tool routing
Manual model selection
No permission visibility
Voice limited to browser

With CoreLayer

Unified control layer
MCP + native tool registry
Smart model gateway
Permission guard + audit logs
Desktop voice pipeline

From request to action, with approval in the loop.

Every tool call follows a structured path through interpretation, routing, and permission verification before execution.

User asks

Through voice, keyboard shortcut, or chat input.

Jarvis interprets

Parses intent, context, and conversation history.

Model Gateway routes

Selects the right model for the task and budget.

Tool Registry resolves

Finds native tools, MCP servers, skills, or REST adapters.

Permission Guard pauses

Risky actions require approval. Audit logs record what happened.

Core capabilities

Six systems working together to give you a desktop AI command layer that respects your permissions and your data.

MCP-first integration

Connect personal apps and external tool servers through MCP.

Unified tool registry

Register, route, execute, and display tools from native modules, MCP, skills, and REST adapters.

Permission Guard

Classify risky actions, pause for confirmation, and keep audit logs.

Model Gateway

Route requests across MiMo, Groq, OpenRouter, local models, and future providers.

Voice pipeline

Wake, listen, transcribe, stream responses, speak back, and support interruption.

Control Center

Manage models, tools, apps, permissions, voice profiles, daemon health, and logs.

Agent actions need boundaries.

CoreLayer treats tool execution as a permissioned system. Risky operations can be classified, paused, approved, denied, and logged.

Permission GuardApproval required

MCP Tool Call

filesystem.write("/documents/report.md")

WRITE
Read
Write
External
Destructive

Agent actions should be visible. CoreLayer classifies tool calls, pauses risky operations for approval, and records what happened so automation stays inspectable.

Voice that belongs to the desktop, not just the browser.

Jarvis supports wake word detection, streaming transcription, TTS with interruption, and a floating overlay designed for desktop workflows.

Wake wordStreaming TTSInterruptionOverlay
Voice Pipeline
CoreLayer voice assistant overlay showing listening state

Built as a desktop control layer.

A modular architecture where each component has a clear responsibility and boundary.

Architecture Overview
CoreLayer architecture from Tauri desktop app to daemon, model gateway, tool registry, permission guard, and storage
Explore architecture

Bring CoreLayer to your desktop.

Start with the desktop app, then connect models, MCP servers, and local tools through the Control Center.

Download latest releaseInstall from source