Skip to content
Projects

huyedits/Symbio has 33 GitHub stars and 2 forks. GitHot measured 0 stars in 24 hours. Last checked 2026-11-15 16:12 UTC. Last push 2026-11-07.

How these counts are measured

Measured star growth

in 24h
0
in 7 days
Collecting history
in 15 days
Collecting history
in 1 month
Collecting history

Net star changes over overlapping periods, not daily totals. Missing measurements are left blank. Stars measure attention, not software quality.

About

Symbio is a self fine tuning AI agent that gives power to users through seamless interaction with their machines. This flexible system adapts to commands, continually learning and evolving based on user input and allows it to learn emperical data.

Machine Learning

README.md

Original

Symbio

A local agent that fine-tunes itself on your corrections — and rolls back the runs that made it worse.

Runs on your Mac. No cloud inference, no subscription, no telemetry by default. What it learns about you goes into weights that only your copy has.

GitHub License

curl -fsSL https://raw.githubusercontent.com/huyedits/Symbio/main/setup.sh | bash

Needs an Apple silicon Mac (M1 or later), 16 GB of memory recommended, and about 8 GB of free disk. Setup picks a model that fits your RAM. Other ways to install ↓


Why

Correct most agents today and they make the same mistake tomorrow. A longer system prompt slows every turn and teaches the model nothing.

Symbio turns corrections into training data. When enough pile up, it runs a small LoRA fine-tune on your machine, checks the result against a regression battery, and throws the new weights away if anything got worse.

you correct it ─► mistake note ─► training sample ─► LoRA fine-tune
                                                          │
          keep the adapter ◄── passes? ◄── golden-set check
                                   │
                                   └─ worse? ─► automatic rollback

What it has measured

Same machine (16 GB M4), held-out tasks, graded by executing the answer or checking the world it changed — not by the model grading itself. Most are one run; treat them as results, not benchmarks.

What Before After
Learning an unfamiliar CLI from its own attempts (curve) 0/13 13/13 by iteration 100
Earning its own training data in a new environment, no human samples (rounds) 10/45 26/45
…six-step tasks in that environment 0/19 7/19
Triggered per-skill adapters vs. no adapter, 60 held-out cases 29/60 59/60
Same 14B, self-repairing 20 injected bugs, harness fixes only 3/20 7/20

And what did not work, measured just as carefully:

  • One always-on adapter for everything scored **51/69 against the base model's 60/69*ne-tune at all. That is why skills get their own adapters.
  • Six-step chains could not be learned from a corpus that contained none. Self-teaching earns vocabulary and short combinations, not arbitrary depth.
  • Training past the first perfect score changed nothing: 900 of 1,000 iterations were

Details, failures included, are further down. Every bench number names its run.

What it does

  • Learns from corrections. It detects them, saves a note, and trains once there a
  • Guards every fine-tune. It runs a golden-set check and rolls back automatically on any regression.
  • Skills with their own adapters. Write a procedure in Markdown, and a small worker model trains on it in about 30 seconds.
  • Decides to train by itself, using its own train_adapter tool. This is logged llback.
  • Uses tools: a sandboxed shell and Python, a browser, desktop control through the accessibility tree, and web research saved to notes.
  • Asks before acting. Sending, posting and deleting need your approval, on any site.
  • Stays yours. Notes, sessions, training data and adapters stay on your Mac: in ~/.symbio for a pip install, or in the checkout for a clone. Upgrading never touches them.
  • Runs where you are: terminal, desktop window, Telegram, or other apps through MCP and ACP.

Honest limits

  • Apple silicon only. MLX has no CUDA path yet; one is written but has never run.
  • A 16 GB Mac fits an 8B comfortably and a 14B tightly. A 14B plus an open browser ca
  • Fine-tunes take minutes and real power: about 17 Wh for 150 steps on the 14B.
  • Self-training can overfit. The rollback catches regressions it can measure, not ones it cannot.
  • The tool sandbox is best effort. It is not a security boundary.
  • This is one person's project, and many features are experimental.

Install

One line

curl -fsSL https://raw.githubusercontent.com/huyedits/Symbio/main/setup.sh | bash

This clones into ~/Symbio (set INSTALL_DIR=... to change it) and runs ./install.sh. Run it again to update.

With pip

pip install "symbio-cli[mlx]"
symb setup
symb chat

To work on it

git clone https://github.com/huyedits/Symbio
cd Symbio
./install.sh --dev

On first launch the setup wizard asks for names, a model and features. Press Enter at the model step to take one sized to your Mac: Qwen3-14B at 24 GB and up, Qwen3-8B from 12 GB, and a 3B below that. Run it again any time with symb setup.

Installer options
./install.sh --prefetch-model   # download the model during install
./install.sh --no-browser       # skip the Chromium download
./install.sh --with-vision      # let the assistant see screenshots (mlx-vlm)
./install.sh --dev              # test dependencies
./install.sh --no-shell         # don't drop into the environment afterwards
./install.sh --venv PATH        # custom virtualenv location