Margot v2.1

A multi-agent AI desktop assistant with 140+ tools and a memory that persists.

A coordinator agent orchestrates specialized subagents across Google Cloud, research, browser automation, and code generation—backed by a four-layer cognitive memory system that persists context between conversations.

Request a Demo
0:00 / 0:00

What it does

Google Cloud

68 tools across Analytics 4, Search Console, Merchant Center, Gmail, Drive, Sheets, and Calendar. Multi-account OAuth with automatic routing.

Research

Three tiers: standard chat (5–15s), light research with subtopic pipelines (30–90s), and deep research with 4 independent agents and cross-pollination (90–180s).

Desktop & Browser

21 desktop tools with three-level shell safety. Browser automation via Playwright with 65 discovered skills and deterministic recovery.

Media Generation

Image generation via Gemini 3 Pro (up to 4K), video via Veo 3.1 and Sora 2, voice synthesis via ElevenLabs with streaming.

Coding Agent

Freya runs on MiniMax-M2 with a RAG-enhanced best practices corpus. Auto-linting, auto-commit, 17× cheaper than comparable models.

Skills & Monitoring

Self-extending tool generation from natural language descriptions (20-skill cap, adjustable in source). Heartbeat-based watch system with cron scheduling, 5 evaluator types, and iOS push notifications. Scheduled task executor runs full coordinator workflows unattended with verification contracts and mutation guardrails.

How it works

A Grok 4.1 Fast coordinator with a 2M token context window runs a ReAct loop (up to 30 iterations) and delegates to specialized subagents. Each agent has its own model, tools, and system prompt.

Research 4× Grok 4.1 Fast, cross-pollination between rounds, 90–180s
Freya MiniMax-M2 (230B MoE), RAG corpus, auto-lint, ~$0.003/task
Browser Gemini 3 Flash loops, 65 skills, deterministic recovery
Tauri 2.0 React FastAPI PostgreSQL + pgvector Tailscale SwiftUI

Memory

Four-layer cognitive memory inspired by Mem0. Memories are extracted after each conversation and assembled in under 150ms.

01

Short-Term

Last 20 messages from the current conversation, kept in PostgreSQL.

02

Semantic

Long-term facts and preferences stored as vectors with HNSW indexing and automatic deduplication.

03

Episodic

Past conversation summaries with importance scoring—top 3 retrieved per query.

04

Procedural

Reusable workflow patterns detected from tool usage sequences with success tracking.

Get Margot

Free and open-source. Bring your own API keys.

~$0.001/query standard ~$0.01 light research ~$0.016 deep research
View on GitHub

macOS 13+ · iOS 17+ companion