Hilbana plugin for Claude Code
Official marketplace for the hilbana plugin for Claude Code:
persistent per-project agent memory, the human+agent workflow framework, and
Hilbana's MCP server auto-registered.
Language note: the commands and skills are written in Spanish. They work
fine whatever language you talk to Claude in — Claude reads them and replies in
yours — but if you plan to edit them, expect Spanish prose inside.
🇪🇸 Español: README.es.md
Install
/plugin marketplace add hilbana/claude-plugin
/plugin install hilbana@hilbana
Restart Claude Code. On install you'll be asked for your Hilbana API key
(format hil_…, created under Settings → API keys in your workspace) and,
optionally, a base URL if you self-host.
Documentation
- REFERENCE.md — every command, skill and hook in detail: what
each one does, its arguments, the MCP tools it uses and the rules it follows. The
source of truth if you're using the plugin day to day.
- Plugin README — what's included, requirements,
verification, token accounting and privacy.
- CHANGELOG — what changed in each version.
Guides, examples and prompts
- Hilbana + Claude Code — the plugin or a plain MCP
connection, how to check it works, your first loop.
(Español)
- Hilbana + Beads — using Beads for an agent's own plan and
Hilbana for the team's board and shared memory, without duplicating work.
(Español)
- Hilbana + engram — run both side by side or move your engram
memory to Hilbana, with the type mapping and an import prompt.
(Español)
- Memory examples — six realistic memories, one per type, with
what makes each good or bad.
- Starter prompts — prompts to paste at the moments that
matter, and a snippet for
AGENTS.md / CLAUDE.md. In English and Spanish.
What's in here
.claude-plugin/marketplace.json the marketplace (a single entry: hilbana)
plugins/hilbana/ the plugin: commands, skills, hooks, scripts, MCP
guides/ Hilbana + Claude Code, Hilbana + Beads, Hilbana + engram (en, es)
examples/memory/ six example memories, one per type
prompts/ starter prompts and an AGENTS.md snippet (en, es)
Hilbana is a hosted product and its code is not open source. This repository holds
what you run on your own machine.
Contributing
Before pushing, run the validator:
node scripts/validate-plugin.mjs
CI runs it on every push and PR, and again as a gate before publishing a release. It
guards the failure mode that bit us once: Claude Code silently drops a component
whose frontmatter doesn't parse, so a stray : in a description can remove a
command with nothing failing anywhere. Hence the rule it enforces — description and
argument-hint are always double-quoted — plus checks on the manifests, the hook
script paths, node --check on the hooks, and REFERENCE.md staying in sync with the
files.
Releases
Versions are published as releases
and recorded in the CHANGELOG. To hear about new ones, set the repo
to Watch → Custom → Releases: Claude Code doesn't notify you on its own, it only
surfaces an update when you open /plugin.
To update:
/plugin marketplace update hilbana
/plugin install hilbana@hilbana
Then restart Claude Code (or run /reload-plugins).
What counts as patch, minor or major
The version field in plugin.json is the source of truth: change it on main and
the release publishes itself.
| Change | Bump |
|---|
| Fix to a command, skill or hook that doesn't change how you use it | patch (1.3.0 → 1.3.1) |
| New command, skill or hook; new opt-in behaviour | minor (1.3.0 → 1.4.0) |
Renaming or removing a command/skill, changing the name or meaning of a userConfig field, or anything that forces users to touch their setup | major (1.3.0 → 2.0.0) |
A plugin has no API to break, but it does have muscle memory: if someone has
/hilbana:plan in their fingers and it disappears, that's a major. That rule is
what made the 2.0.0 release a major: the cycle commands moved to the MCP server.
Privacy
The plugin contains no secrets. Your API key lives only in your local Claude Code
configuration (userConfig, marked sensitive) and is never published here.
License
MIT
