---
slug: "skillopt"
name: "SkillOpt"
effect: "Trains a skill document for a frozen LLM agent with validation-gated edits, shipped as best_skill.md"
category: "devtool"
tier: "rising"
url: "https://aka.ms/skillopt"
github: "microsoft/SkillOpt"
pricing: "open-source"
license: "MIT"
tags: ["skills", "skill-optimization", "self-evolving-agents", "validation-gate", "python"]
platforms: ["claude-code", "codex"]
install_kind: "pip"
date_added: "2026-10-18"
rank: 19
rising_score: 48
stars: 18520
star_velocity: 113.5
velocity_basis: "since creation"
created_at: "2026-05-08T06:41:01Z"
pushed_at: "2026-10-17T19:30:59Z"
signals_status: "ok"
signals_fetched_at: "2026-10-18T11:41:04.466Z"
page: "https://agenttooldrop.com/tools/skillopt"
---

# SkillOpt

Trains a skill document for a frozen LLM agent with validation-gated edits, shipped as best_skill.md

Category: Devtools. Tier: new and rising, rank 19 by new and rising score (48 of 100). Pricing: Open source. License: MIT.

Platforms: claude-code, codex.

## Why you would want it

SkillOpt treats the skill document as the trainable state of a frozen agent and trains it without touching model weights. A separate optimizer model turns scored rollouts into bounded add, delete or replace edits, and in the default path an edit is accepted only when it strictly improves a held-out validation score. The deployed artifact is a best_skill.md file that runs against the unchanged target model, and the README says it adds zero inference-time model calls at deployment. A skillopt-sleep CLI, labelled preview, applies the same gate to past coding-agent sessions and stages proposed updates for you to adopt.

## Best for

- Training a skill file against scored tasks
- Tuning skills for Codex or Claude Code runs
- Reviewing past coding-agent sessions into skills (preview)

## Use it when

You have tasks you can score and a held-out split, and you want a skill document for an agent whose model you keep frozen improved by measured edits instead of hand-tuning.

## Skip it if

You have no scored tasks or validation split to gate edits against, or no access to a configured model backend for the rollout and optimizer calls. The README calls SkillOpt-Sleep a preview and some multi-objective, replay and dream-rollout controls experimental, and the docs say its interfaces and defaults may change. With a real backend, Sleep sends truncated excerpts of harvested session transcripts to the provider you pick, and the docs say outbound prompts are not guaranteed secret-free, so review transcripts first.

## Editor note

PyPI serves 0.2.0. The docs say some newer features, such as Cursor and Pi support and the openai_compatible research backend, need a source install from main, and the benchmark configs and agent integration files live in the repository rather than the wheel.

## Install

Install from PyPI:

```sh
python -m pip install skillopt
```

Source: https://github.com/microsoft/SkillOpt/blob/main/docs/guide/installation.md#pypi

## Live signals

Fetched 2026-10-18T11:41:04.466Z (status: ok).

- Repository: microsoft/SkillOpt
- Repository created: 2026-05-08
- Stars: 18,520 (GitHub's count; it cannot be authenticated)
- Stars per day, since repo creation: 113.5 (an average over the repo's life, not a growth rate)
- Forks: 1,760
- Open issues and PRs: 56
- License (GitHub): MIT
- Last push: 2026-10-17
- Latest release: v0.2.0 (2026-07-02)

## Links

- Website: https://aka.ms/skillopt
- GitHub: https://github.com/microsoft/SkillOpt
- Install instructions verified at: https://github.com/microsoft/SkillOpt/blob/main/docs/guide/installation.md#pypi
- Source for this entry: https://github.com/microsoft/SkillOpt#readme
- AgentToolDrop page: https://agenttooldrop.com/tools/skillopt
- Full catalog (JSON): https://agenttooldrop.com/tools.json