opchain
skills that ship.
If you've ever tried to build something real with Claude and watched it lose the plot between chats — forgetting your spec, redesigning what it already designed, repeating questions it asked an hour ago — opchain is what you've been looking for.
A set of skills plus a checkpoint protocol — native in Claude Code, and now in Codex or any MCP agent via a hosted endpoint. Each skill specializes in one piece of the puzzle (discovery, spec, design, build, audit, ship, monitor, migrate, scale). At resume time, the saved checkpoint helps you pick up where you left off. Ask the orchestrator at any point and it reads that state back to you and names the skill to run next.
Install the skills and call them by name. Each starts with a readable Markdown file. Shared commands include their runtime files. Optional task tests use your chosen provider and credentials.
Keep what worked. Test what should change.
Opchain 2.0 brings progress, lessons and proposed rules together. You can review the evidence before changing how a skill works.
It remembers.
Import scored outcomes into local history. Duplicate records count once, and source references stay attached.
It proposes.
Three weak results can flag a pattern. A proposed rule must improve the target task without breaking other cases.
You decide.
Activation requires a signature from an externally trusted reviewer. Lessons and rules expire and remain advisory.
Concept → shipped, in one chat
Idea → deployed, one chat
resources
References and trust signals you'd otherwise have to scroll all the way to the footer to find.
Skills
Read the skill files, references and runtime code. You can inspect, fork and change them.
read the files ›Glossary
Plain-English definitions for every opchain term — checkpoint, tri-agent, audit gate.
read ›How it's built
opchain.dev is itself built with opchain. Every commit ran through the gates.
read ›Security
Threat model, dependency posture, secret handling — what we actually do.
read ›