Skip to content

Get started

Get started · Beta

Install Trimtab.

One command installs the beta on your Mac. Then run trimtab and pick a training project.

Step 1: Requirements

Apple Silicon Mac
M1 or newer. Intel Macs, Linux and Windows are not supported in this beta yet.
A model provider
An API key for one provider (Anthropic, OpenAI, Gemini, …), or a local model with Ollama or LM Studio.
Nothing else to install
The installer brings its own Python 3.11 and never touches yours (conda, venv, Homebrew).

Your training script

  • Takes the number of epochs as a command-line flag (default --epochs).
  • Exposes the settings you want tuned as command-line flags; a script with none is refused, not guessed at.
  • Reports a metric every epoch: with from trimtab import report, a printed TRIMTAB_METRIC line, or a Lightning metrics.csv.

Step 2: Install

Paste this into Terminal:

curl -fsSL https://raw.githubusercontent.com/inferenceflowai/trimtab-beta/main/install.sh | bash

Installs uv (a small tool) if you don’t have it, lets it fetch a private Python 3.11, creates an isolated environment at ~/.trimtab/app, downloads Trimtab from the GitHub release, checks its SHA256 and links trimtab into ~/.local/bin. Running it again upgrades in place. Your data in ~/.trimtab is never deleted.

View repository on GitHub The installer, releases and README live in the beta repository.

Step 3: Launch

Open a new terminal window if the installer says so, then run:

trimtab

Trimtab opens in your browser at http://127.0.0.1:8765. On the first start:

  1. A short Welcome: exactly what is sent to the AI provider and what stays on your Mac.
  2. The beta terms: click I agree.
  3. Open Settings in the left rail and add your provider’s API key. Keys stay on your Mac.

trimtab: command not found? Open a new terminal window (the installer added ~/.local/bin to your PATH), or run ~/.local/bin/trimtab.

Step 4: First run

  1. Select your training project

    Choose the folder. Trimtab finds the trainer, the Python environment and the metric, and shows what it detected. No project handy? Try the bundled CIFAR-10 or quickstart example.

  2. Configure the search length

    Quick (3 ideas, about 5 minutes), Standard or Thorough. The estimated time and token use show before you start.

  3. Choose a model

    Pick the provider and model in the picker at the end of the card, if you want a different one.

  4. Start the search

    Press Start search and watch the trials and the chart come in. Stop at any time.

  5. Review the recommendation

    Read the changes and the evidence, then copy the command for the full GPU run.

Step 5: Updates and support

Updating
Run trimtab update in a terminal. The app also tells you in its footer when a new version is out.
Uninstalling
Run trimtab uninstall. It asks before deleting your data.
Feedback
Use Feedback in the app’s left rail, or “Did this help?” under a recommendation. You see the exact text before it is sent.

Known limitations

macOS only in this beta
Apple Silicon Macs.
Local models are experimental
Small Ollama or LM Studio models can miss the reply format; a hosted model is more reliable.
Hard-coded hyperparameters can’t be tuned
A script with no command-line options is refused instead of guessed at.
Tested on 5 projects so far
A CIFAR-10 demo, BERT4Rec, a Lightning example, a Hugging Face fine-tune and a Keras script it refused.
Laptop results are a proxy
The command was verified on a Mac, not on a multi-GPU run, so check the first GPU epochs.
A search costs LLM tokens
The estimate shows before you start and the real count after.
Data your trainer downloads stays
Trimtab lists what appeared in your project and doesn’t delete it.

Documentation and terms