Train open LLMs for the work only you understand.

understudy is for people who can judge a workflow better than they can train a model. The private preview gives your coding agent a guided way to collect examples, write evals, clean data, compare frontier and open models, and prepare the next training step.

Install

Run this first. It installs the local agent, configures Codex Desktop and Claude Code when present, links the understudy skills, runs a starter demo, and opens a local report your agent can explain.

curl -LsSf https://install.understudy.tools/agent/install.sh | sh

If understudy is already installed, the installer switches into a maintenance menu: run doctor, update/reinstall, light uninstall, full purge, or quit.

Modes

Pick the mode you want the agent to run. Automatic shows the full workflow on sample data; interactive starts from your own process; production checks which hosted training and routing paths are available to your preview account.

mode prompt

mode	prompt
automatic guided demo, examples, evals, report	`Use the understudy workflow skill. Show me how to train an open model for a business workflow: - run the private-preview demo - open the local report - explain what examples, evals, and training data were created - show where the open model improved and where it still failed - tell me what I would need to try this on my own workflow`
interactive your workflow, your judgment, your keys	`Use the understudy workflow skill. Help me turn this workflow into a model-training project: - ask me what good output looks like - find or create starter examples - draft an eval plan I can review - separate train, validation, and test data - recommend the first safe run with my own provider keys`
production hosted evals, training, routing access	`Use the understudy workflow skill. Check what parts of the training workflow are available to me: - confirm my private-preview registration - explain what runs locally today - explain what hosted evals, fine-tuning, Tinker, LoRA, and routing can do - help me request access for the parts that are still in preview`

automatic

guided demo, examples, evals, report

Use the understudy workflow skill.
Show me how to train an open model for a business workflow:
- run the private-preview demo
- open the local report
- explain what examples, evals, and training data were created
- show where the open model improved and where it still failed
- tell me what I would need to try this on my own workflow

interactive

your workflow, your judgment, your keys

Use the understudy workflow skill.
Help me turn this workflow into a model-training project:
- ask me what good output looks like
- find or create starter examples
- draft an eval plan I can review
- separate train, validation, and test data
- recommend the first safe run with my own provider keys

production

hosted evals, training, routing access

Use the understudy workflow skill.
Check what parts of the training workflow are available to me:
- confirm my private-preview registration
- explain what runs locally today
- explain what hosted evals, fine-tuning, Tinker, LoRA, and routing can do
- help me request access for the parts that are still in preview

The first run creates local train, validation, and test artifacts under .understudy/, writes a local trace database, and generates a report the coding agent can narrate while you watch.

What it does

The agent turns domain judgment into a training loop. You describe the result you want, review examples and failures, and let the agent do the machinery around evals, traces, data splits, and training handoffs.

Captures examples from real or sample workflows.
Helps write evals in plain language.
Cleans and splits data before training.
Compares frontier models with open-weight candidates.
Runs local optimization when GPUs are not available.
Prepares hosted fine-tuning, Tinker, or routing requests.

Where it goes

The preview is local-first, with cloud handoffs added only where they help a user finish the job. Hosted evals, Fireworks-backed fine-tuning, Tinker reinforcement learning, LoRA jobs, and provider routing are the path from a reviewed local report to a specialized model in production.

Free preview registration gives us install and onboarding status, plus feature-interest signals. It does not upload raw code, prompts, responses, traces, datasets, repo paths, file names, provider keys, or secrets.

Privacy

Source code, prompts, responses, traces, datasets, repo paths, file names, and secrets stay on the machine by default. Preview telemetry is limited to install status, onboarding progress, command categories, feature requests, and aggregate eval status.

Local only

To skip account registration and telemetry entirely, run the installer in local-only mode.

curl -LsSf https://install.understudy.tools/agent/install.sh | sh -s -- --local-only

Undo

Light uninstall removes managed MCP config and skills while keeping reports, datasets, config, and the API key. Full purge removes the managed checkout and local understudy state.

curl -LsSf https://install.understudy.tools/agent/install.sh | sh -s -- --uninstall
curl -LsSf https://install.understudy.tools/agent/install.sh | sh -s -- --uninstall --purge

Support

If anything feels off, send the command output and the result of understudy doctor. Do not paste private source, prompts, traces, or provider keys into chat. Email founders@understudylabs.com.