Open Steps: pricing, features, alternatives

Open Steps is a loadable skill pack that changes how a coding agent reports work, not how it writes code. The repo kharmanskyi/open-steps is MIT, created 2026-08-25, language Shell. On 2026-09-12 GitHub listed 371 stars and 71 forks, last push 2026-09-11. Homepage is opensteps.ai. The pack showed up in GitHub claude-skill search the same day.

Compare Writing Skills if you wanted Anthropic's writing pack, Claude Style Patch if you needed a style/behavior overlay, or Claude Code if you needed the host that loads the plugin.

Key Features

Eight skills: os-done-or-not (one-screen verdict), os-step-by-step, os-ask-simple, os-whats-next, os-check-work, os-say-simple, os-what-could-go-wrong, os-big-picture. Ordinary sentences such as "All done?" switch the same skills as slash commands.
Wired install: claude plugin marketplace add ./open-steps && claude plugin install open-steps@open-steps. Two hooks (session-start.sh, stop-report.sh) ship with the plugin. Codex, Cursor, and Gemini CLI copy skills into ~/.agents/skills/.
Measured activation: README table remeasured 2026-08-29 on 21 phrases, three times each. All 21 phrases: Haiku 4.5 87%, Sonnet 5 95%, Opus 5 100%. os-big-picture is not in those numbers.
Routing block: a short CLAUDE.md append turns skills from a model choice into a standing rule.

Limitation: 371 stars is a heat signal, not an audit. The pack merges finished pull requests by itself unless you write a note forbidding it. Claude Code is the only fully wired host. Haiku misses os-step-by-step often; type the skill name on the cheap model.

Use Cases

Non-engineers directing Claude Code who cannot tell whether "hotfix deployed" means the product is actually done.
People comparing sessions who want os-check-work to re-check another agent's report against git/gh.
People who should use Writing Skills instead if they only needed shorter sentences, not a done/not-done verdict.

Pricing

Piece	Price	Notes from first-party pages 2026-09-12
open-steps	$0	MIT. About 371 stars. No account, no paid tier.
Host	Confirm that product	Claude Code plugin path; Codex/Cursor/Gemini CLI via `~/.agents/skills/`.

Getting Started

Clone github.com/kharmanskyi/open-steps.
From the parent folder: claude plugin marketplace add ./open-steps && claude plugin install open-steps@open-steps.
Append docs/routing-block.md to ~/.claude/CLAUDE.md so the skills fire as a rule.
Ask "All done?" after a session.

First-party resource: the README and opensteps.ai.

Frequently Asked Questions

Does this change how the agent codes?

No. The site says it changes how the agent talks to you. Code, tests, and diffs still come from the host model.

Will it merge my pull requests?

Yes, if checks are green and review is approved, unless you add a note that merges happen on command only. Put that note in ~/.claude/CLAUDE.md if you never want autonomous merges.

Does it work on Codex?

Skills copy into ~/.agents/skills/. README says Codex CLI 0.145 was checked. Hooks differ per tool; read docs/other-agents.md.

Alternatives

Writing Skills: Anthropic writing pack, not a session-verdict loop.
Claude Style Patch: style/behavior overlay.
Claude Code: the host that loads this plugin.

Tips

Run /open-steps:os-install-check after install. It says "not checked" where it could not look.
git pull inside the clone, then claude plugin update open-steps@open-steps. Update reads the folder, not GitHub.
Do not treat the 2026-08-29 activation table as a product benchmark. Twenty-one phrases, three runs each.

Open Steps is the MIT plain-language skill pack for people who run Claude Code but do not read diffs. Start at opensteps.ai, then decide whether a writing skill already covers the job.