Every skill is a single SKILL.md file. Install the full set or just the ones you want. No API keys, no package manager, no build step.
01
Claude Code CLI
Need the CLI first?Install Claude Code, then pick a path. The plugin is the one to use if you can — it carries the skills and the hooks that enforce them. The zip carries the skills only.
Plugin recommended
Two commands. Adds the 29 skills, eight registered slash commands, and three hooks: a commit gate that refuses git commit unless the code being committed passed bug-check, pipeline state at session start, and a pointer to the next skill when one finishes.
Drop the skills into .claude/skills/; Claude Code discovers them next session. Use this where a plugin install isn't an option. Note what it does not include: none of the hooks, so nothing mechanically enforces a gate — the skills describe the contract and you follow it.
Verify it worked. In that session, type/oc-discover. If oc-app-architect starts an interview ("Who's the primary user? Sync or async? …"), you're set. If not, re-check that the files landed in.claude/skills/<name>/SKILL.md.
Start a new conversation and name the skill ("use oc-app-architect — I have an app idea") or use its slash command (/oc-discover). Naming it explicitly is markedly more reliable than hoping the description matches.
03
Team (check into git)
Commit the skills directory so every teammate uses the same prompt pipeline.
Codex — and any MCP-aware agent (Claude Desktop, Cursor, Windsurf) — reaches opchain two ways, over theModel Context Protocolor as drop-in skills. The drop-in path is the closest match to Claude Code: Codex Agent Skills auto-trigger by description, so the pipeline fires on its own. The hosted MCP endpoint needs no install, but relies on the model choosing to call opchain's tools.
Drop the skills in (recommended) — Codex Agent Skills use the same SKILL.md format and auto-trigger by description — the same unenforced, name-it-yourself matching Claude Code skills rely on:
Verify it worked. Run codexand ask "build me an app" — a drop-in skill match (or, over the hosted server, opchain's route tool) hands you to oc-app-architect's discovery interview. Note: Codex doesn't surface MCP prompts, so opchain's /oc-* verbs aren't slash commands in Codex — just describe what you want ("build me an app", "audit this") and the matching skill triggers.
upgrade
Update an existing install
Already running opchain in a project? Replace the skills in place. Your .checkpoints/ directory livesoutside.claude/skills/, so every skill's session state — phase, step, next actions, decisions — is preserved across the swap. The new skills pick up the existing checkpoints on their next invocation.
Claude Code CLI
Run from your project root. This wipes the old skills directory and replaces it with the latest bundle.
# .checkpoints/ is a sibling of .claude/ — never touched by this swap
rm -rf .claude/skills/
curl -L https://opchain.dev/opchain-skills.zip -o opchain-skills.zip
unzip opchain-skills.zip -d .claude/skills/
rm opchain-skills.zip
Verify the new version is live. Check what the site is currently shipping withcurl https://opchain.dev/api/health (theversion field is the build SHA), then start a new Claude Code session and run/oc-ops — the orchestrator should report your existing checkpoints exactly where you left them.
Claude.ai / Claude Desktop
Download the latest bundle ZIP or the individual SKILL.md files you use from the Skill Library.
Open Claude → Settings → Customize → Skills.
For each skill you've installed, click the existing entry and choose Replace (or delete and re-upload). Skills are matched by name, so the new file overrides the old one.
Checkpoints in Claude.ai live in the project, not the skill, so they survive the upload.
Team (git-tracked install)
If you committed .claude/skills/ with the team flow, do the upgrade on a branch so the skill diff lands in review:
Returns the build SHA currently published atopchain.dev. Re-run after upgrading and confirm a fresh Claude session sees the new skills under.claude/skills/<name>/SKILL.md.
Get notified automatically
Skills don't phone home, so a stale install stays stale silently. This Claude Code hook checks at session start: it reads theversion from your installed skills' frontmatter, compares it against the live catalog'scatalogVersion atopchain.dev/skills.json, and drops a one-line notice into the session when a newer release is out. Merge it into ~/.claude/settings.json (or a project's .claude/settings.json) — if you already have a hooks block, add theSessionStart entry to it rather than pasting the whole object.
It checks, it never installs — releases can carry breaking changes (see the changelog), so the swap stays a deliberate step: run the upgrade command above when you see the notice. When you're current, offline, or opchain isn't installed, the hook prints nothing and gets out of the way. It fires only onstartup — not on resume or compact — and gives the network three seconds before giving up.
Configuration
.claude/skills/
Where Claude Code looks for skills.
.checkpoints/
Skills write JSON session state here. Add to .gitignore.
orchestrator.md
Shared welcome protocol + pipeline map. Bundled with every skill.
.codex/skills/
Where Codex looks for drop-in skills — same SKILL.md format as Claude Code.
~/.codex/config.toml
Register the hosted MCP server under [mcp_servers.opchain].
Session persistence — JSON checkpoints in .checkpoints/; pick up exactly where you left off, next chat or next week
Documented handoffs — each skill names the one that follows it, so you always know what to run next; with the plugin installed, the commit gate is enforced and finished skills suggest the next step
Natural-language triggers — type "build me an app" or "audit this before Friday" and opchain routes you to the right skill
Want a preview first? The demo walks you through twelve end-to-end opchain sessions.
what next
Your first 30 minutes
Skills are inert until you trigger one. Here are three starter paths — pick whichever matches the moment.