I've been waiting for a model that understands sound. Seed Audio is the first time TTS felt over

Marwane Manifi — Wed, 01 Jul 2026 04:47:50 +0000

I make AI videos, and for the longest time the sound was the part I quietly dreaded.

Not because a clip has no audio. Because every voice tool I tried just didn't understand the scene. They could read a script fine, but the second a character needed to actually feel something, the voice went flat and robotic. The ambience never matched the shot. I'd still end up hunting for effects and mixing the whole thing by hand to make thirty seconds feel like one real moment. The visuals kept
getting better every month. The audio stayed stuck at "good enough."

Then last week I tried Seed Audio, and it kind of stopped me. It's the first one that feels like it actually understands what it's making.

What it is

Seed Audio is ByteDance's new audio model. You tell it what you want to hear, a line, a whole scene, a mood, and it generates the entire thing together. Voices with real emotion, several characters talking to each other, the room they're in, the sound effects around them, all in sync. Not a bare voice track you have to score and patch afterward. The actual moment.

If you remember the point where image models suddenly understood pictures instead of guessing at them, this is that same jump, for sound.

It does far more than one thing

That's the part that's hard to get across in text. The same model will do a calm narration, a two-person argument, a meditation that keeps getting slower and softer, a stand-up bit where it adds its own room laughter, a tense ghost story, cinematic scenes, weather, crowds, whole environments. Different feel every time,
and each one sounds like it belongs to the scene instead of being pasted on top of it.

voice reference

You can hand it a short clip of a voice and it locks onto that voice, the tone, the pacing, the little habits in how someone talks. Then you reuse that exact voice anywhere.

For video this is the thing I've wanted forever: one consistent character voice across a whole series, from a single reference, without re-recording anything. I cloned a voice once and dropped it into completely different scenes, calm, panicked, tired, amused, and it held up in all of them. Every cloning tool I'd tried before nailed exactly one read. This one carried.

voice a character from a picture

Here's the one that really got me as a video person: you can hand it an image of a character and it works out a voice that fits the face. My characters usually start as a still, so pulling a voice straight from that image, no casting, no guessing, feels almost unfair.

What you can actually make with it

you can use seed audio to handle all kinds of scenarios.

Film & video voiceover — narration and character voices that match the mood of a shot instead of sitting flatly on top of it.
Sound effects & ambience — rain, wind, crowds, room tone, whole environments, without digging through a stock library.
Podcasts — multi-host conversations that actually sound like two people talking, not one voice split in two.
Audio drama & multi-character scenes — short dramas and dialogue-heavy scenes cast and performed in a single pass.
Audiobooks & narration — long reads that hold one character's voice consistently from the first chapter to the last.
Ads & brand spots — a finished voiceover-and-effects spot straight from a single description.

It's a 1.0, so it's not flawless. Sometimes a read comes back a little flat and I regenerate, and it's strongest in English and Chinese right now. Small stuff. Didn't change how I feel about it. AIGenTools helped me complete a full day's work in just one hour, giving me more time to do other things.

If audio has always been the weak link in your AI videos, also you can try seed audio，You will also discover some aspects that truly amaze you.

Hermes Agent: The Self-Improving AI Agent That Runs 24/7 on Your Own Server — A Deep Dive

Marwane Manifi — Tue, 14 Apr 2026 13:34:30 +0000

What Is Hermes Agent?

Hermes Agent is an open-source AI agent framework built by Nous Research (the team behind the Hermes model family). It runs on your own machine or server, connects to any LLM you choose, and can be reached through Telegram, Discord, Slack, WhatsApp, Signal, and more — 14 platforms total.

But what makes it genuinely different from the dozens of other agent tools out there is this: Hermes builds its own harness and improves itself over time. Every time you use it, it gets a little better at working with you specifically.

It's not a chatbot. It's not a code completion tool. It's an autonomous agent that can run shell commands, read files, browse the web, schedule tasks, delegate to sub-agents, and remember everything about how you work — across sessions, across platforms, across months of use.

The Five Core Mechanisms

1. The Learning Loop — It Gets Better Without You Teaching It

This is the heart of Hermes. After every task, it runs an automatic retrospective:

What should I remember? → writes to memory
Is this a recurring pattern? → creates a Skill
Did the user correct me? → updates the Skill

You don't trigger this. It just happens. The first time you ask it to write a Python script, the output might be generic. By the tenth time, it knows you prefer httpx over requests, you like error logs written to files instead of printed to terminal, and you hate long functions. Nobody taught it any of that — it observed and extracted the patterns itself.

2. Three-Layer Memory — From Goldfish to Old Friend

Most AI tools forget you the moment you close the tab. Hermes has three memory layers:

Session memory (SQLite + FTS5): Raw conversation history, full-text indexed. Hermes doesn't dump all of this into context — it searches by topic and pulls only what's relevant. Three months of use feels as fast as three days.
Persistent memory: A small, curated profile — your preferences, habits, tool choices. This loads into every conversation automatically.
Skill memory: Markdown files that capture "how to do things" — your validated workflows, stored in ~/.hermes/skills/.

The key design choice: on-demand retrieval, not full-context loading. FTS5 full-text search means Hermes finds the right memory at the right time without bloating token usage.

Optional add-on: Honcho user modeling (by Plastic Labs) — a 12-layer dialectical identity system that infers things about you that you never explicitly said. Like noticing you always pick the cheapest option and inferring you're cost-sensitive, so next time it surfaces pricing info first.

3. Self-Evolving Skills

A Skill in Hermes is a standalone markdown file in ~/.hermes/skills/ — instructions for how to do something well. Skills come from three sources:

Source	How it works
Bundled	40+ pre-built Skills ship with installation
Agent-created	Hermes auto-creates Skills from recurring tasks
Community Hub	Install with one command

The killer feature: Skills self-improve through use. Every time you give feedback ("next time check if the table exists first"), Hermes goes back and edits the Skill file. Next run, the improvement is baked in.

Skills follow the agentskills.io standard — they're portable across Claude Code, Cursor, Codex CLI, Gemini CLI, and more. A Skill you wrote for one tool works in another.

4. 40+ Built-in Tools + MCP

Five categories of native tools:

Category	Examples	What they do
Execution	terminal, code_execution, file	Run commands (sandboxed), read/write files
Information	web, browser, session_search	Search, scrape, search conversation history
Media	vision, image_gen, tts	Understand images, generate images, text-to-speech
Memory	memory, skills, todo, cronjob	Manage memory, Skills, scheduled tasks
Coordination	delegation, moa, clarify	Sub-agents, multi-model answering, ask user

Beyond the built-in tools, Hermes supports MCP (Model Context Protocol) — connecting to 6,000+ external apps. GitHub, Slack, Jira, databases, Google Drive — add a config block to config.yaml and it just works. You can also filter which tools each MCP server exposes for security.

5. Multi-Platform Gateway

One unified process connects to all your messaging platforms simultaneously. Start a conversation on Telegram during your commute, continue it in the terminal at your desk — Hermes doesn't distinguish where a message came from. Same brain, same memory, same Skills.

Supports: Telegram, Discord, Slack, WhatsApp, Signal, Email, SMS (Twilio), Home Assistant, Mattermost, Matrix, DingTalk, Feishu/Lark, WeCom, and Open WebUI.

How It Compares: Hermes vs Claude Code vs OpenClaw

These three tools aren't competing — they solve different problems.

Dimension	Claude Code	OpenClaw	Hermes Agent
Core philosophy	Interactive coding	Configuration-as-behavior	Autonomous background + self-improvement
Your role	Sitting at the terminal directing	Writing config files	Deploy and check in occasionally
Memory	CLAUDE.md + auto-memory	SOUL.md + Daily Logs + semantic search	Three-layer self-improving memory
Skill source	Manually installed	ClawHub 5,700+ community	Agent-created + community Hub
Run mode	On-demand	On-demand	24/7 background
Deployment	Local CLI (subscription)	Local CLI (free + API costs)	Self-hosted (VPS / Docker / Serverless)

The practical way to think about it:

Need real-time pair programming? → Claude Code
Need transparent, auditable agent behavior for a team? → OpenClaw
Need 24/7 automated code review, scheduled tasks, persistent memory? → Hermes
Long-term content project? → Hermes for research + Claude Code for writing

They're complementary. The agentskills.io standard means Skills are portable between all three.

Getting Started: From Zero to Running in 15 Minutes

Step 1: Install

curl -fsSL https://raw.githubusercontent.com/NousResearch/hermes-agent/main/scripts/install.sh | bash
source ~/.bashrc
hermes version  # verify

Works on macOS, Linux, WSL2, and even Termux on Android. The script handles Python, Node.js, and all dependencies.

Docker alternative:

docker pull nousresearch/hermes-agent:latest
docker run -v ~/.hermes:/opt/data nousresearch/hermes-agent:latest

Step 2: Configure a model

All configuration lives in one file: ~/.hermes/config.yaml

model:
  provider: openrouter       # or nousportal, openai, anthropic, ollama, etc.
  api_key: sk-or-xxxxx
  model: anthropic/claude-sonnet-4
terminal: local              # or docker, ssh, daytona, modal

Supported providers include OpenRouter (200+ models with one key), Nous Portal (officially recommended), OpenAI, Anthropic, z.ai/Zhipu, and Ollama for fully offline local models.

Step 3: Start talking

hermes

That's it. Ask it to search the web, read files, run shell commands. It will ask for permission before executing anything potentially dangerous.

Step 4: Connect Telegram (optional but game-changing)

Create a bot with @botfather in Telegram → get a token
Add to ~/.hermes/config.yaml:

gateway:
  telegram:
    token: YOUR_BOT_TOKEN

Run hermes — it connects to the Telegram Gateway automatically

You now have a personal AI assistant reachable from your phone, anywhere in the world.

Step 5: Deploy to a VPS for 24/7 uptime

Any cheap VPS will do (Hetzner, DigitalOcean, Vultr — whatever you already use). Hermes is lightweight without a local LLM.

docker run -d \
  --name hermes \
  --restart unless-stopped \
  -v ~/.hermes:/opt/data \
  nousresearch/hermes-agent:latest

That's a 24/7 AI assistant with persistent memory, reachable from your phone.

What Makes It Actually Special

After spending a week with this, here's what I think the real differentiators are:

The harness builds itself. Other tools need you to write the rules (CLAUDE.md, SOUL.md). Hermes observes, learns, and writes its own rules. You can override anything, but the default is zero-config improvement.
Memory that actually works long-term. FTS5 on-demand retrieval means it doesn't slow down after months. Your data stays local in SQLite — no cloud vector database needed.
Everything stays on your machine. MIT license, self-hosted, any LLM you want. No vendor lock-in, no data leaving your server. A basic VPS runs it comfortably — memory usage is under 500MB without a local model.
Skills are portable. The agentskills.io standard means your investment in Skills isn't locked to one platform. Switch tools, keep your workflows.
It works when you're not there. Cron scheduling + Telegram gateway = an agent that reviews PRs at 2am, monitors your servers, drafts daily reports, and sends you the results over breakfast.

Limitations to Be Aware Of

No automatic memory expiration. The memory database only grows. You need to manually audit ~/.hermes/ occasionally.
Skill self-improvement depends on feedback quality. Vague feedback ("this doesn't feel right") won't improve anything. Specific feedback works.
Auto-generated rules can drift. The convenience of automatic improvement comes with the trade-off that you're giving up some control. Check your Skill files periodically.
Community ecosystem is still young. OpenClaw's ClawHub has 5,700+ Skills from the lobster craze's network effects. Hermes's community is growing but isn't there yet.

DEV Community: Marwane Manifi