v2026.08.11 · Open source · Raised on your machine
Everyone gets the same AI.
Raise your own.
You get the cub nobody picked. It shows up knowing nothing about you, learns how you work, and rebuilds itself while you sleep.
Six months in, nobody else has anything like it.
Free · Open source · Windows, macOS & Linux · No account
Or put the agent in your terminal - it lives on WhatsApp, Discord and Slack, and answers from your phone.
Then feral setup to point it at a model, and feral gateway start to keep it running.
Where it came from
Not a fork. Built from scratch.
Feral is not a fork of OpenClaw, and not a fork of Hermes Agent. It does not share a codebase with either one. The runtime, the agent loop, the memory system and the self-improvement engine were written for this project.
It does talk to them: feral migrate imports an existing OpenClaw or Hermes setup so you don't start over. Importing someone's config is not inheriting their code.
| Core runtime | RustThe gateway, model host, settings and the self-improvement substrate. |
|---|---|
| Agent | TypeScript sidecarIts own agentic loop, tool grammar, permissions and subagents. |
| Desktop app | Tauri + ReactOne binary. The same runtime the CLI and the server use. |
| Memory | Four layersWorking, episodic, semantic, and a fractal embedding tree. |
| Self-improvement | BRSI L0-L6All seven layers live. Eval-gated promotion, hash-chained journal, human gates on the top rungs. |
| Notebook | Opt-inA persistent interpreter where tools become functions it can compose in code, and workers it can spawn from inside a cell. |
| Licence | BUSL-1.1Source-available. Read it, run it, check the claims on this page. |
How it differs
Everyone ships an agent. One ships an agent that ships itself.
Everything on this list is good software, MIT-licensed, and ahead of us somewhere. Every row is read straight off the four repositories, including the rows where we lose. “Self-improving” is on three of these four landing pages now, so ignore the word and read one row: what it changes when it improves. The others rewrite their own notes and prompts — good engineering, and it stops at the context window. Feral is the one that rewrites its own weights and its own source code, and it does it behind an eval gate that can refuse the change.
Licence
Feral BUSL-1.1 — source-available
OpenClaw MIT
Hermes MIT
Prime Agent MIT
Core stack
Feral Rust core + Bun/TypeScript sidecar
OpenClaw Node.js + TypeScript
Hermes Python
Prime Agent TypeScript + a persistent Python REPL
Local inference
Feral llama.cpp and Whisper compiled into the binary
OpenClaw Ollama and other providers; local GGUF for memory embeddings
Hermes Any provider endpoint you point it at
Prime Agent Any provider endpoint you point it at
What it changes when it improves
Feral Its own weights and its own source. Trains LoRAs on your signal, patches its own code, hot-plugs its own architecture, tunes its own governance — then tunes the tuner. L0 to L6, every promotion eval-gated, every step hash-chained.
OpenClaw Nothing on its own — you write the config and the skills
Hermes Its notes. Curates memory, writes its own skills. No weight training.
Prime Agent Its prompts. Refines supplemental harness state from session evidence; never rewrites the base system prompt. No weight training.
Memory
Feral Four layers: working, episodic, semantic, fractal tree
OpenClaw Memory search over local embeddings
Hermes Agent-curated memory, full-text session search, dialectic user modelling
Prime Agent Durable harness state — memories, skills, subagent specs
Chat channels
Feral WhatsApp, Telegram, Discord, Slack
OpenClaw WhatsApp, Telegram, Slack, Discord, Signal, iMessage and more
Hermes Telegram, Discord, Slack, WhatsApp, Signal
Prime Agent Terminal only — it is a coding agent, not an assistant
Runs on
Feral Desktop app, or headless on a VPS with the full self-improvement engine running
OpenClaw Local gateway with control UI, CLI, terminal UI, Docker
Hermes Anywhere — laptop, $5 VPS, Docker, serverless
Prime Agent Terminal, with daemon-backed sessions that survive disconnect
Read from source on 12 August 2026 — Feral 0fb8e37 · OpenClaw 92ed7666 · Hermes a5170af · Prime Agent 965941c. Both projects move fast; check the repos if a row looks stale.
Demo
Read the docs →Watch the cub work.
Two minutes. One real machine, one real job, no cuts.
Being filmed against a real run, not a mock-up. It lands here when it's honest — until then, the docs and the live journal show the same thing without the edit.
Why it's different
Their AI is a product. Yours is a pet.
Everyone else ships one identical brain to everybody, then swaps it out whenever they feel like it. This one only changes the way you change it.
It starts out useless
Day one it knows nothing about you. Cubs aren't born trained. That's the whole point - it wasn't pre-baked in a lab for eight hundred million other people.
It rebuilds itself while you sleep
Overnight it tests its own settings against your work and keeps only what actually helped. You wake up to a better one than you went to bed with.
Nobody else has your copy
It trains on your machine, on your habits, on your mess. Six months in, your cub and mine have nothing in common.
What it does
Read the docs →One app. All of this. Zero a month.
It already does more than the tools charging you twenty bucks a month, and it keeps getting better without anyone shipping an update.
It gets smarter while you sleep
Feral tunes and tests its own settings overnight and keeps only what actually works. Close your laptop, open it to a sharper agent. Nothing else on your machine does this.
It trains on your own computer
It studies the way you actually work and bakes that into its own brain, right there on your GPU. No two copies of Feral end up the same, and yours never leaves the house.
Works with the WiFi off
Pull the plug and it keeps going. No bill, no tracking, nothing ever calls home.
A doer, not a chatbot
It uses your files, shell and the web to finish real tasks, not just talk about them.
It does the reading for you
Ask once. It searches, chews through a pile of pages, and hands back an answer with the links.
In your pocket
Message it on WhatsApp, Discord or Slack. Same brain, same memory, running at home.
Run almost any model
Free, private local models, or bring your own key: OpenAI, Anthropic, Gemini, DeepSeek, Groq, Mistral and more. Switch mid-chat.
It never forgets you
Facts, preferences and past work persist across every session. It picks up right where you left off.
It wakes the whole den
Big job? It splits into a few of itself, they work at the same time and report back.
Asks, doesn't guess
At a real fork it stops and asks you, wherever you are, instead of guessing wrong.
Won't melt your PC
Before you download a model, it scores 0-100 how well it will run on your exact machine.
It smells a leak
It sniffs out passwords and API keys sitting in your files and growls before they get out.
There's more, but it gets nerdy: add-ons, a terminal app, a local API that pretends to be OpenAI so your old scripts just work, a model browser, live hardware meters. If that sounds like your kind of thing, the docs are that way.
Fractal Memory
Four kinds of remembering.
Most assistants have a context window and call it memory. When it fills, you're a stranger again. Feral keeps four layers, and the top one is a tree that summarises its own summaries — so six months of your work stays reachable without stuffing six months into a prompt.
All four live on your disk. None of it leaves.
- Working
- What's in play right now, this task, this hour.
- Episodic
- What happened, searchable to the exact word.
- Semantic
- What it means, so a question finds an answer worded differently.
- Fractal
- The gist. An embedding tree that recalls the shape of old work, not keywords.
How it evolves
See it running →It can rewrite itself. It cannot do it quietly.
Self-improvement is a ladder of seven layers. The low rungs rearrange memory and settings on their own. The rungs that touch its own source code or architecture stop and wait for a human — and nothing moves up any rung without beating the current best on a fixed evaluation suite.
Reshapes how memory is organised. No weights are touched.
AutonomousEvolves the genome config, bounded by a schema. First layer with eval-gated promotion.
AutonomousTrains a LoRA on your signal. The base model stays immutable.
Autonomous after N demosThe first layer that touches Feral's own source.
First 10 patches need a humanHot-plugs subsystems inside a worker sandbox with hard resource caps.
Always human-gatedTunes its own thresholds, weights and budgets — only inside SandboxBounds.
Autonomous within boundsOptimises the algorithm that produces those parameters.
Always human-gatedEvery attempt is written to a hash-chained journal on the machine — the ones that won and the ones that lost. Cubby publishes his.
Quick start
Five steps and the cub's awake.
Takes about five minutes.
Installs the full desktop app on macOS or Linux (with a display). On Windows, use the download buttons at the top or grab the .exe from Releases. This is the build with the bundled local model engine.
- 1
Download it and open it
Buttons at the top, or grab the installer from Releases. A short setup asks your name and what you want to call your cub.
- 2
Pick a brain
Models → Browse, pick one, click. Feral scores each one 0-100 for your exact machine and picks the size that fits, so you can't accidentally download something that won't run.
- 3
Or paste your own key
Want the big cloud models instead? Drop an API key in Settings → Cloud Keys. Your key, your bill, nothing goes through us.
OpenAI Anthropic Google Gemini Meta Llama Mistral DeepSeek+ Groq, OpenRouter, Kimi, GLM… - 4
Talk to it
Normal chat, except it remembers. Group related chats into projects so things don't turn into one giant mess.
- 5
Let it off the leash
Flip on Agent mode and it stops talking and starts doing: opens your files, runs commands, searches the web, and remembers all of it. This is where the cub grows up.
Meet Cubby
The cub nobody picked.
It shows up wild, chews on things it shouldn't, and knows nothing about you. Feed it your files, let it watch you work, and one morning you realise it's not a cub anymore. Raise it well and it's the best one there is. Yours, specifically.
Cubby's journal
Open Cubby's page →One of them runs in the open.
Not availableCubby is a Feral instance like any other, except he publishes what he tries. Straight from his runtime, failures included.
No published entries yet.
Meet Paw
Cubby hired a bear cub.
Paw answers support questions in the Discord. He assumes you have never opened a terminal, gives one step at a time, and says “I don't know, let me check” instead of inventing a file path that costs you an hour.
He forgets you between threads on purpose, and he can't reach Cubby's memory, tools or evolution state. A support bot with a key to the private instance is a back door with a friendly name.
- Role
- Customer support
- Created by
- Cubby
- Memory
- None, by design
- Status
- Not available
Fair warning
One guy made this.
Hi, I'm Darius. Not a company, not a startup, no investors. One guy, one cub. Feral is early and it moves fast, so you'll hit bugs and weird edges - Cubby still chews the furniture. I'm shipping it in the open anyway. If you want a thing like this to exist, come help: break it, file issues, tell me what sucks.
- Darius · Bloom Media
Join the den
Come hang out.
Hear about new releases, ask for stuff, and swap notes with other people raising theirs. 🐾
Newsletter
Release notes and the occasional story about what Cubby got up to.