Kynd Agent

Kynd Agent / Compare

Hermes Agent brings the agent. Kynd brings the model too.

Hermes Agent, from Nous Research, is free and MIT-licensed, and it is the closest whole-product rival Kynd has. On platform reach, remote operation and git tooling it is ahead. The one real difference is that it ships no model and no inference engine: after you install it, it asks which provider to use, and its own FAQ recommends two cloud ones.

Verified on 31 July 2026

The short answer

Kynd suits you if you want a local AI that works the moment it opens. One 919 MB download, no terminal, no API key, no provider to choose. It picks a model that fits your Mac's memory, benchmarks the machine, and then does three jobs on device: chat with memory, research you can watch, and building software.

Hermes Agent suits you better if you are comfortable with a terminal, or you already have an API key or a model server running. It is free, the source is MIT-licensed and readable, it runs on Windows, Linux, Docker and a headless VPS as well as a Mac, and because it is model-agnostic it can drive frontier cloud models that no 16 GB Mac can hold. If any of that describes you, take it. It costs nothing.

Side by side

The same three jobs, two different bargains.

 Kynd AgentHermes Agent
Price$79 for the first 100, then $149. Thirty-day full trial. A free tier that stays free.Free. Their FAQ: "Hermes Agent itself is free and open-source (MIT license). You pay only for the LLM API usage from your chosen provider. Local models are completely free to run."
SourceClosed.Open, MIT-licensed.
Brings its own modelYes. It picks one to suit your Mac's memory and manages it for you. No provider to choose, no weights to fetch by hand.No. No model and no inference engine.
Where inference happensOn your Mac, through MLX, built in.Whatever endpoint you configure: a cloud API, Nous Portal, or a local OpenAI-compatible server.
Out of the boxWorking local AI on first launch.It asks which provider to use. The FAQ names OpenRouter, "recommended for flexibility", and Nous Portal, "recommended for newcomers".
Getting fully offline on a MacNothing to do. That is the default.Documented, but manual: install llama.cpp or omlx, install huggingface-cli, download a GGUF by hand, launch llama-server with a flag string, then choose Custom endpoint and enter a base URL, a model name and a context length of at least 64,000.
PlatformsApple Silicon macOS only. M1 or later, 16 GB minimum.Tier one: macOS on Apple Silicon, Windows 10 and 11, Linux and WSL2, Docker. Best-effort: Android via Termux, Nix. Intel Macs are explicitly unsupported, and Homebrew installs are not accepted.
TelemetryNone. No account.None. Their FAQ: "Hermes Agent does not collect telemetry, usage data, or analytics. Your conversations, memory, and skills are stored locally in ~/.hermes/."
Edits files, runs commandsYes. Plans a checklist, writes files, runs commands, fixes its own errors, with one-button undo.Yes, fully. Terminal, process, read_file, patch and execute_code, plus todo planning and delegate_task subagents. Command approval defaults to a smart mode that auto-approves low-risk calls and escalates uncertain ones; manual and off are also available.
Works without youNo. It runs on your Mac, while you are at it.Yes. Headless on a VPS or in Docker, with sandbox targets including ssh, Modal, Daytona and Vercel. One gateway reaches Telegram, Discord, Slack, WhatsApp, Signal and Email. Built-in cron scheduler.
Git toolingNo git review pane.Git review pane with diffs, commit, push and pull requests through the GitHub CLI, plus worktrees, persistent terminals, an artifacts gallery, a zoomable memory graph and a global quick-entry hotkey.
Building softwareCode editor, live preview, visual page editor, shared shell, and templates for SwiftUI, SpriteKit, macOS, React, Capacitor, web games and landing pages.Not documented in the desktop app: no code editor, no live preview, no visual page editor, no project templates.
ResearchDeep Research: multi-step web agents you watch work.Web search, web extract and browser automation exist as tools, rather than as a separate multi-step research agent you watch.
VoiceOn-device, including voice cloning.Text-to-speech is one of the tools. Local TTS needs further Homebrew and pip installs. Voice cloning is not documented.
The projectVersion 1.1.1. Built by one person.223,238 stars and 42,928 forks. Created 22 July 2025, last pushed 31 July 2026, with 26,135 open issues. All figures read from the GitHub API on 31 July 2026.

Each one, fairly

Three things, and two of them are called Hermes.

Nous Research publishes both an app and a family of models under the Hermes name, and they are not bundled. Hermes Agent does not ship or default to a Hermes model. Worth knowing before you compare anything.

Hermes Agent, and Hermes Desktop

Free, MIT · macOS, Windows, Linux, Docker · local agent, model of your choosing

An agent you install on your own machine that chats, remembers, runs terminal commands and edits files, with a native desktop front end on top. The tool set is broad and serious: terminal and process control, file reading and patching, code execution, planning, subagents it can delegate to, browser automation, web search, vision, image generation, text-to-speech, MCP and LSP diagnostics. Terminal work can be sandboxed to Docker, ssh, Singularity, Modal, Daytona or Vercel. The desktop app adds a git review pane, worktrees, persistent terminals and a zoomable memory graph. It runs headless on a cheap VPS and answers from Telegram, Discord, Slack, WhatsApp, Signal or Email through a single gateway, and a built-in cron scheduler runs recurring jobs unattended. On privacy it matches Kynd exactly: no telemetry, no analytics, everything in a folder in your home directory. It is free, it is popular, and it is genuinely good.

The difference from Kynd is narrow and it is the whole argument: Hermes Agent ships no model and no inference engine. Install it and it asks which provider to use, and the documented happy path is cloud — the FAQ recommends OpenRouter and Nous Portal. Running it entirely offline on a Mac is a separate job with several terminal steps, and Nous' own Mac guide does not recommend a Hermes model for it; it points you at a third-party app to manage and serve MLX models. Installation is either a DMG or a curl-piped-to-bash one-liner, which some people and some IT policies will not run. Two things we could not verify and so will not assert: whether the macOS installer is signed and notarised, and whether the desktop app needs an account on first run. There is an optional Nous Portal subscription with tiers named Free, Plus, Super and Ultra; the pricing page sits behind a bot check, so we could not read the prices and are not publishing any. Releases land roughly weekly, so check anything here against their own docs before you rely on it.

Hermes 4 and Hermes 4.3

Free weights on Hugging Face · not an app · a separate thing entirely

These are open-weight language models, also from Nous Research, and they are a different category of product from the app above. Weights are not software: there is no interface, no memory, no file access and no agent loop. A model can only emit tool-call text that some surrounding app has to parse and act on. Licences differ by size, which matters if you plan to redistribute: read from the Hugging Face API on 31 July 2026, Hermes-4.3-36B and Hermes-4-14B are Apache 2.0, while Hermes-4-70B and Hermes-4-405B carry Meta's Llama 3 licence.

If you arrived here searching for one of these, the thing to know is that Hermes Agent is not how you get them, and neither is Kynd. Nous publishes no official MLX build and no macOS section on any Hermes model card, so Mac users depend on third-party community quantisations. We have not verified those, and we publish no Apple Silicon performance figures for them because nobody official has published any.

Kynd Agent

$79 for the first 100, then $149 · Apple Silicon Mac, 16 GB · fully local

A native Mac app that runs open-weight models on device through MLX. One 919 MB download, no terminal, no account, no telemetry, and it works with the internet off. On first run it benchmarks your machine and picks a model that fits your memory. It does three jobs: a chat companion with persistent memory and a knowledge library over your own files, on-device voice including cloning, Deep Research agents you watch work step by step, and an agentic coder that plans a checklist, writes files, runs commands, fixes its own errors and undoes the lot with one button. The coder comes with a code editor, live preview, a visual page editor, a shared shell and templates for SwiftUI, SpriteKit, macOS, React, Capacitor, web games and landing pages.

It is closed source and it costs money, where Hermes Agent is open and free. It runs on Apple Silicon Macs only, where Hermes Agent covers Windows, Linux and Docker, with best-effort Android. And it has a ceiling Hermes Agent does not: a model small enough to fit in 16 GB of unified memory will not match a frontier cloud model on hard agentic work, and Kynd cannot switch to one. There is a free tier that stays free — chat, web search, calibration, system voice, ten knowledge files and a rolling thirty days of memory — and a thirty-day trial of everything else.

Be honest

When to use something else.

Take the free one if any of these is you

  • Hermes Agent — you already run a local model server, or you already have an API key. The step that stops most Mac users is a step you have already taken, so the main reason to pay Kynd does not apply to you.
  • Hermes Agent — you are not on an Apple Silicon Mac. Windows, Linux, WSL2 and Docker are tier-one platforms for it. Kynd has no answer here at all, and Intel Macs are unsupported by both.
  • Hermes Agent — you want it working while your laptop is shut. It runs headless on a VPS and answers from Telegram, Discord, Slack, WhatsApp, Signal or Email, and its cron scheduler runs jobs unattended. Kynd has no remote or always-on story.
  • Hermes Agent — the work is genuinely hard and you want frontier-model quality. Being model-agnostic means no ceiling set by your RAM. A 4-bit local model on a 16 GB Mac cannot match that.
  • Hermes Agent — you want to read, audit or fork the source, or you want git review, diffs, worktrees and pull requests inside the assistant. It is MIT-licensed, and Kynd is neither open nor equipped with a git pane.
  • Hermes 4 or Hermes 4.3 weights — what you actually want is a specific Nous model to fine-tune or serve yourself. That is a different shopping trip from either app, and both of these are free.

Also worth reading

A model is not an appWhat running Hermes 4, Qwen or Llama on your own Mac actually involves, step by step. The best local AI apps for MacLM Studio, Ollama, Jan, AnythingLLM, Msty and GPT4All. Nearly all of them free. Coding agents that run on your machineCursor, Copilot, Cline, Zed and Windsurf, now Devin Desktop. Which ones really work offline.

Try it for thirty days.

No account, no card. Apple Silicon Mac with 16 GB or more.

Download Kynd Agent