blog

opchain speaks MCP now

opchain grew up inside Claude Code — skills on disk, read by one agent, in one CLI. As of v1.4.3, that’s no longer the boundary: there’s a hosted MCP server at opchain.dev/mcp, and any MCP-speaking agent — Codex first among them — can reach the skill catalog over the wire. A point release with a headline feature, which deserves both the announcement and an explanation of the version number.

What shipped

  • The hosted server. JSON-RPC over streamable HTTP at https://opchain.dev/mcp. It serves the same catalog the skill library renders — same source of truth, different audience. No API key, no signup; it’s a docs surface that answers tool calls.
  • The Codex install flow. /install now has a path for Codex and generic MCP clients alongside the Claude Code one. Point your client at the endpoint, and the skills that were “a Claude Code thing” become “a thing your agent can consult.”
  • The plumbing honesty. The Worker that serves this site now also terminates MCP. One deploy ships both, which keeps the server honest: it physically cannot describe a catalog other than the one the site ships.

The version number is a confession

Sharp-eyed roadmap readers will note the MCP server was slated for v1.8 — several releases out. We built it early anyway, sat on it briefly, and shipped it as a point release. The honest sequence: the prototype turned out to be a weekend, not a quarter, because the catalog already had one machine-readable source of truth (the same generated file the site builds from). When the expensive part of a roadmap item evaporates, holding the feature hostage to its original quarter is process cosplay. So: 1.4.3, “extend opchain to Codex,” months ahead of its own plan. We’ll take the scheduling embarrassment; it points the right direction.

Why this matters more than a port

The narrow read is “opchain works in another tool now.” The wider one is about what a skill is. On disk, a skill is instructions an agent follows. Over MCP, the same skill is queryable expertise — an agent mid-task can ask “how does the opchain pipeline gate a deploy?” and get the actual contract, current as of the last deploy, instead of a months-stale training impression of our docs.

That reframes the catalog from product for Claude Code users to reference any agent can consult. The pipeline discipline — spec gates, evaluator loops, checkpoints — stops being locked to one runtime. Agents are becoming each other’s users; serving them well is not a novelty feature, it’s distribution.

Try it

Claude Code users: nothing changes, the skills stay local-first — install as ever. Codex (or anything MCP-speaking): add https://opchain.dev/mcp and ask it what opchain would do before you merge that branch. The answer comes from the same catalog, whichever door you use — one product, now with two runtimes and, as of this week, no favorite child.

The opchain team

Builders of opchain

We build opchain — a skillchain and checkpoint protocol for shipping real software with Claude. We write about what we learn dogfooding it on our own pipeline.

← All posts