RUDI Registry

Official registry of MCP stacks, binaries, agents, runtimes, and skills for the RUDI CLI. The catalog uses schema version 2 with unversioned canonical paths.

Package Types

Type	Description	Canonical source
Stack	MCP servers with tools	catalog/stacks/{id}/manifest.json
Binary	Standalone binaries and CLIs	catalog/binaries/{id}.json
Agent	AI coding assistants	catalog/agents/{id}.json
Runtime	Language interpreters	catalog/runtimes/{id}.json
Skill	Reusable agent workflows	catalog/skills/{id}/SKILL.md

Workers compose goals, processes, stacks and skills, context, operating boundaries, and completion criteria. The workers authoring area currently contains manually loaded Markdown definitions and a synthetic Client Meeting Update example; it is not an installable package type.

index.json is the single generated package index. Do not add parallel version-suffixed files or directories; the schema version belongs inside the document, not in its path.

Usage

rudi search whisper
rudi install whisper
rudi install ffmpeg
rudi list

Repository Structure

index.json                    # Generated schema-v2 package index
catalog/
├── stacks/{id}/
│   ├── manifest.json         # Canonical stack metadata
│   └── src/, node/, python/  # Stack implementation
├── skills/                   # Same-ID skill folders
├── workers/                  # Markdown worker definitions; authoring only
├── binaries/{id}.json        # Canonical binary metadata
├── agents/{id}.json          # Canonical agent metadata
└── runtimes/{id}.json        # Canonical runtime metadata
dist/                         # Generated indexes, hashes, and release metadata

Edit catalog manifests and skill sources by hand, then regenerate the index:

npm run indexes:sync
npm run indexes:check

Creating a Stack

1. Create catalog/stacks/{stack-id}/.
2. Add catalog/stacks/{stack-id}/manifest.json.
3. Add the MCP implementation and focused tests.
4. Run the required checks below.

Minimal manifest:

{
  "id": "stack:my-stack",
  "kind": "stack",
  "name": "My Stack",
  "version": "1.0.0",
  "lifecycle": {
    "maturity": "experimental",
    "support": "supported"
  },
  "delivery": "remote",
  "install": {
    "source": "catalog",
    "path": "catalog/stacks/my-stack"
  },
  "runtime": "node",
  "surface": "local-only",
  "requires": {
    "binaries": [],
    "secrets": [
      { "key": "MY_API_KEY", "label": "API key", "required": true }
    ]
  },
  "provides": { "tools": ["my_tool"] },
  "related": {
    "operatorSkill": "skill:my-stack",
    "skills": ["skill:my-stack"]
  },
  "mcp": {
    "transport": "stdio",
    "command": "node",
    "args": ["src/index.js"]
  },
  "meta": {
    "description": "What the stack does",
    "author": "Your Name",
    "license": "MIT",
    "category": "data",
    "tags": ["example", "capability:query"]
  }
}

Every stack must have one primary operator skill. Add its standalone catalog skill, declare the stack in that skill's requires.stacks, set related.operatorSkill, and include the same ID in related.skills. Additional related.skills entries are optional companion workflows. provides.tools is only for MCP tools exposed by the stack.

Stacks share the seven primitive categories and capability/domain/provider tags with their primary operators. See Stack catalog organization for naming, discovery, verification and reproducible index checks.

surface classifies hosted eligibility as local-only, cloud-hosted, or both; omission fails closed to local-only. A both stack uses exact toolSurfaces overrides to opt individual tools into hosted execution. See SCHEMA.md for the no-elevation and allowlist rules.

lifecycle is optional and omission means unclassified. Do not claim stable or supported without evidence. Deprecated packages carry an announcement date and migration message; retired packages are removed from the canonical catalog rather than left as installable placeholders. See SCHEMA.md.

Creating a Skill

Author each skill in catalog/skills/{skill-id}/SKILL.md. Legacy flat files remain readable for compatibility. Bundles may include scripts/, references/, and assets/. Catalog packages must be portable: personal paths, client state, account data, and brand-specific defaults belong in local/private skills or ~/.rudi state.

An operator skill translates user intent into actual stack tool calls. It must name its required stack in frontmatter, use the live MCP tool schema as the parameter authority, define mutation confirmation and failure behavior, and verify externally visible changes when the stack supports a read-back.

Adding a Binary, Runtime, or Agent

Create one JSON manifest in the matching unversioned directory. Downloaded artifacts require a pinned version, HTTPS URL, supported extraction type, and a SHA-256 checksum for every platform. See SCHEMA.md for the complete contract.

Discovering Inventory

Do not maintain a hardcoded package list in documentation. Generate the current inventory from the canonical index:

node -e "const i=require('./index.json'); console.log(Object.keys(i.packages).join('\\n'))"

Required Checks

npm test
npm run validate
npm run indexes:sync
npm run indexes:check
npm run catalog:clean:check
npm run stacks:verify -- --changed-from origin/main --prepare
npm run debt:scan
npm run build
npm run release:verify
npm pack --dry-run --json

URLs

• Index: https://raw.githubusercontent.com/learnrudi/registry/main/index.json
• Catalog stacks: https://raw.githubusercontent.com/learnrudi/registry/main/catalog/stacks/{id}/
• Binary releases: https://github.com/learnrudi/registry/releases/download/{tag}/{name}

Security

Never commit API keys, credentials, tokens, account state, downloaded media, or runtime output. Stacks declare secret names in manifest.json; secret values are stored locally through RUDI and are never part of the registry.

See CONTRIBUTING.md for contribution and isolated-install guidance.

License

MIT

Skill category, naming, facet and dependency conventions are defined in Skill catalog organization.