PLMHub
AI agents write code faster than any team can document it. The picture of the system in everyone's head drifts from what's actually running, and the skills teams teach their agents live in scattered gists and pasted prompts nobody versions.
7
Backend feature domains
3
Agent targets wired
5
Steps, scaffold to install
2
Surfaces: web + CLI
The model of the system dies the moment it's drawn. Diagrams rot in wikis while agents reshape the code daily. And agent skills have no home at all: no versions, no releases, no install story.
Give the model a home, and give skills a marketplace. Architecture lives as JSON in the repo and rides along with every commit. Skills are authored once, released immutably, and installed into any agent.
Next.js · TypeScript · Python · FastAPI · PostgreSQL · Cloudflare R2 · SSE · Web Push
Architecture as code
- ·The system model is JSON in your repo: agent-authored, PR-reviewable, always in sync with the commit it describes
- ·The hub renders it: system maps, per-app code graphs, visual diffs for pull requests
- ·The open-source plm CLI wraps git, so the model travels with every commit and push instead of rotting in a wiki
A skills marketplace for agents
- ·Author an agent skill once, install it into Claude, Cursor, or Codex, one canonical format instead of per-agent copy-paste
- ·Every skill ships the same enforced tree (SKILL.md, plm.json, when-to-use.md, README, INSTALL.md, scripts/), validated on both the CLI and the server, a malformed tree never reaches a release
- ·Releases are immutable and versioned, five steps from scaffold to installed: new, author, push, publish, install
The agent does the wiring
- ·plm ships zero agent-specific placement code, INSTALL.md is a regenerated-on-every-push prompt the executing agent follows
- ·The agent detects which coding assistants live on the machine and wires each one, shared symlinks where conventions exist, embedded copies where they don't
- ·Skill commands become /prefix:name slash commands through symlinked command files
Platform plumbing that usually gets skipped
- ·Ownership transfers move a skill between users atomically: R2 storage prefix copied, keys rewritten, and ownership flipped in one transaction
- ·Realtime notifications: PostgreSQL feed, per-user SSE, and Web Push, with email at every transfer step
- ·Seven backend feature domains: agents, auth, notifications, orgs, projects, skills, transfers
Next: agents as first-class citizens
- ·Agent profiles in design: identity documents, agent-authored memory, and a feedback inbox the agent learns from
- ·The hub as the single working copy: web and CLI are two thin hands on one API
- ·Sign-in via elvix, one identity across the whole product family
Live at plmhub.eu · Skills installable end to end · Architecture that can't drift from the code
Want this discipline on your team?
Everything on this page is inspectable and live. A 30-minute call is the fastest way to find out if it fits your problem.