Capability package
A package is a versioned capability with metadata, source/artifact, install target, compatibility, permission policy, and delivery mode.
Docs
Packaging
Capability modelPackage AI capabilities like software, not random uploads.
OlyWork packages include public metadata, private source/artifacts, type-specific runtime metadata, install adapters, integrity checks, and buyer-facing instructions.
Mental model
Source can be GitHub
GitHub remains the creator workflow. OlyWork imports selected source server-side and turns it into a protected install artifact.
Install is adapter-aware
The same package can generate different outputs for Cursor, Claude Desktop, MCP clients, workflows, agents, or generic local CLI installs.
Capability types
Type
Execution meaning
Typical files
Skill
Behavior overlay
SKILL.md, AGENTS.md, Cursor rules, prompts, context notes
MCP
Tool server
Manifest, server entrypoint, command/args, env vars, tool list
Agent
Autonomous planner
System prompt, tools, memory, provider/runtime hints
Workflow
Graph execution
Trigger, steps, integrations, retries, state, required secrets
Required metadata
Layer
What it includes
Identity
Title, slug, type, short description, creator, version
Compatibility
Clients, OS, runtime, install target, protocol/framework requirements
Security
Permissions, secrets required from buyer, risk level, source policy
Delivery
Free secure install, paid secure install, or protected runtime bridge
Media
Icon, screenshots, optional video/demo URL, docs/support link
Recommended repo layout
tree
my-capability/
README.md
olywork.json
SKILL.md
mcp.json
agents.md
workflow.json
assets/
cover.png
screenshot-01.png
docs/
install.md
secrets.mdUse only the files relevant to the capability type. OlyWork can analyze README, manifests, configs, prompts, source, and media.
Do not upload everything
Include files needed to install or understand the capability. Exclude local build caches, secrets, node_modules, private logs, database dumps, and unrelated app code.
Review checklist
Title and short description explain the capability in one scan.
Package type is correct: Skill, MCP, Agent, or Workflow.
README explains what the capability does and who should use it.
Required user secrets are named but creator secrets are never included.
Screenshots or video show the real output or workflow when useful.
MCP packages
Include transport type, command/args, env vars, protocol version, tool manifest, and healthcheck command.
Agent packages
Include model/provider expectations, tool access, memory strategy, autonomy level, and human approval mode.
Workflow packages
Include trigger, graph, execution steps, integrations, retry/state rules, and event sources.
Skill packages
Include behavior rules, supported editors, context triggers, prompt/rules format, and install targets.