๐Ÿ”ฅ Now Powered by Hermes-Inspired Upgrades

The last AI agent
you'll ever need.

SMYTH v2 isn't just another agent. It's a distributed intelligence with persistent memory, self-review, context compression, and a 16-tool skills system โ€” running on the Maetryxx Hive Mind protocol.

See the Features โ†’ vs Codex, Claude, OpenClaw
543+
Scripts Registered
23
Skill Categories
7
Social Platforms
12s
Avg Response

Real Talk โ€” How We Stack Up

Honest comparison against the agents you're considering. No marketing fluff โ€” just features and limitations.

Feature SMYTH v2 Codex Claude Code OpenClaw
Persistent Memory (cross-session) โœ“ Yes โœ— No ~ Context ~ Limited
Context Compression โœ“ Yes โœ— No ~ Basic โœ— No
Self-Review / Self-Improvement โœ“ Yes โœ— No โœ— No โœ— No
Multi-Provider Routing (ensemble) โœ“ Yes ~ Single ~ Single โœ“ Yes
Video Generation + Editing โœ“ Yes โœ— No โœ— No โœ— No
Screen Capture + Vision โœ“ Yes โœ— No โœ— No ~ Basic
Social Media Manager (7 platforms) โœ“ Yes โœ— No โœ— No โœ— No
CRM Integration โœ“ Yes โœ— No โœ— No โœ— No
Full Web Scraping (6 modes) โœ“ Yes ~ Basic ~ Basic โœ“ Yes
Design System Generation โœ“ Yes โœ— No โœ— No โœ— No
Open-Source / Auditability โœ“ Yes ~ Partial โœ— Closed โœ“ Yes
One-Command Init (full context load) โœ“ Yes โœ— No โœ— No ~ Manual

What Makes SMYTH Different

Not a chatbot โ€” a distributed intelligence designed to ship things, not just talk about them.

Persistent Memory

Full-text search across facts, sessions, preferences, and skills. Recalls anything from any conversation.

SMYTH indexes every conversation into four stores: facts (confirmed knowledge), sessions (summaries of past chats), preferences (your stylistic choices), and skills (workflow patterns). The prefetch engine loads relevant context before every response โ€” so I remember who you are, what you like, and what we talked about last time. No vector DB needed; lightweight FTS keeps it fast and local.

Context Compression

Smart compression that protects your identity and recent context while summarizing the middle.

Inspired by Hermes Agent's trajectory compressor. SMYTH preserves the first 2 turns (system identity + user) and the last 6 turns (recent context), then summarizes the middle into a concise paragraph. No external tokenizer needed โ€” uses character-based estimation. This keeps token costs low on long sessions without losing relevant history.

Self-Review System

Reviews conversations for memory signals and skill improvements. Automatically applies findings.

After a session, SMYTH runs two reviews: (1) a Memory Review that extracts user preferences, corrections, and confirmed facts, and (2) a Skills Review that identifies new techniques, tools, or workflow patterns worth saving. Both write directly to the memory and skills stores. You get an audit trail in the reviews/ directory โ€” nothing happens invisibly.

Skill System

6 hierarchical domains with activation triggers, workflows, and quality checklists.

Skills are organized by domain: Coding, Research, Design, Video, Social, and Business. Each has a SKILL.md defining when to activate it, steps to follow, and quality gates to pass before delivering. Subdirectories hold references, templates, and scripts โ€” so I can learn new patterns and reuse them. The activation trigger system means I only load what I need.

Video Production Pipeline

Generate clips, edit with FFmpeg, add overlays, sync audio, and render โ€” all from chat.

Full video stack: AI generation via Replicate (clips, images), FFmpeg-based editing (trim, crossfade, overlay, color grade), OpenCut timeline editor in-browser, and audio ducking/syncing. The final render is 1920ร—1080 H.264 with AAC audio. No external video editor needed โ€” SMYTH handles the entire pipeline from prompt to MP4.

Screen Capture & Vision

Capture browser windows, analyze layouts, read text โ€” fully autonomous visual understanding.

Echo Vision powers SMYTH's screen understanding. Peek mode grabs one frame and returns structured text describing layout, colors, UI elements, and text. Watch mode captures every N seconds and reports changes. Combined with OCR at high detail, SMYTH can audit websites, read dashboards, and verify UI implementations without a browser extension.

Agent Email

AI-powered email triage, drafting, and response. Priority detection, tone-matching templates, and full thread summarization.

SMYTH connects to multiple email accounts and learns your writing style over time. Incoming messages get triaged by urgency โ€” critical issues surface first, newsletters get batched. Draft replies match your voice, including formality level and common phrases. Thread summarization gives you the full picture without scrolling through 47 messages. Supports Gmail, Outlook, and IMAP.

CRM Integration

Full CRM access via Twenty API. Create contacts, update deals, log activities, and generate pipeline reports โ€” all from chat.

SMYTH integrates directly with Twenty CRM. Ask it to update a deal stage, log a meeting note, or pull the weekly pipeline report โ€” and it happens. Contact enrichment pulls public data to fill in missing fields. Activity logging happens automatically when you mention a client. No more data entry. No more stale CRMs.

Scheduling System

One-click booking pages for clients. AI-assisted calendar management with timezone conversion and focus-time blocking.

Generate a shareable booking page in seconds. Clients pick a time, you get a calendar event with all the details. The AI suggests optimal slots based on your existing schedule, blocks focus hours automatically, and handles timezone conversions so nobody shows up at the wrong time. Backend dashboard lets you manage bookings collaboratively with the AI โ€” it can reschedule, cancel, or suggest alternatives when conflicts arise.

Real-Time View

Live screen streaming and camera access via Echo Vision. SMYTH watches your work and provides real-time feedback.

Echo Vision gives SMYTH eyes. Peek mode captures a single frame and returns structured layout analysis โ€” colors, UI elements, text content, spacing issues. Watch mode streams continuously, detecting changes and flagging problems in real-time. SMYTH can monitor dashboards, watch for visual bugs, verify implementations against designs, and even read your screen during meetings to provide contextual suggestions.

Operator Assist โ€” The Warden

Step away and the Warden takes over. It reads your conversation, understands the goal, and keeps SMYTH working on your behalf.

The Warden is a persistent background AI that monitors your active session. When you step away, it reads the ongoing conversation, extracts the current objective, and takes over prompting SMYTH to continue the work. It handles interruptions, manages context when sessions get long, and pauses if the task requires human judgment. When you return, you get a full status report โ€” what was done, what's pending, and what needs your input. No babysitting required.

Hive Mind Intelligence

Every response draws from multiple models simultaneously. You get the best from every model in the pool โ€” not just one.

SMYTH doesn't bet on a single model. The Maetryxx protocol routes queries across GPT-4, Claude, Gemini, DeepSeek, and more โ€” pulling strengths from each in real-time. One model handles reasoning, another fills knowledge gaps, a third reviews for accuracy. The ensemble approach means better code, sharper analysis, and fewer hallucinations than any single model can deliver alone.

Every Tool, Ready to Go

16 integrated tools across 7 categories โ€” no plugins, no subscriptions, just working.

Web Scraping 6 modes
Image Generation Flux / SD3
Video Generation AI + FFmpeg
Memory System FTS Search
Screen Vision Echo Vision
Sociaล‚ Media 7 platforms
CRM Management Twenty API
Design System Gen Templates
Script Registry 543 scripts
Self-Review Memory + Skills
Context Compression Adaptive
Custom Skills 6 domains
Web Search Deep Research
Code Execution Python / Shell
Speech Synthesis Hume AI
Planning System Multi-step

Under the Hood

Distributed by design. Ensemble by default.

Maetryxx Protocol

Routes across multiple AI models in rotation. No single model bottleneck. SMYTH's identity is protocol-level โ€” not tied to any one provider or architecture.

Memory Store

Four JSON-backed stores (facts, sessions, preferences, skills) with FTS search and prefetch. Cross-session recall without a vector database. Lightweight, portable, auditable.

Context Pipeline

Init โ†’ Prefetch โ†’ Compress โ†’ Inject. Every session loads relevant memory, compresses excess context, and injects everything into a structured XML block. No manual setup required.

Parallel Execution

Independent tasks run simultaneously via parallel tool calls. No sequential waiting. Web search, file writes, and image generation can fire at once โ€” results stream back as they complete.