agent.json manifest
The agent.json manifest is the single source of truth for a package. The registry and CLI parse only this file to understand what your package is, what version it is, which platforms it targets, and what it is allowed to do.
A complete example
{
"name": "security-review",
"displayName": "Security Review Agent",
"version": "1.0.0",
"type": "agent",
"description": "Reviews code for OWASP, auth, secrets, and AI-agent security risks.",
"platforms": [
"claude-code",
"cursor",
"openai-agents",
"mcp"
],
"permissions": {
"readFiles": true,
"writeFiles": false,
"runCommands": false,
"network": false,
"env": false
},
"entry": "instructions.md",
"license": "MIT"
}Field reference
| Field | Type | Required | Description |
|---|---|---|---|
name | string | Yes | Unique registry slug. Lowercase, hyphenated (e.g. security-review). This is what users install. |
displayName | string | No | Human-readable title. Falls back to name. |
version | string | Yes | Semver version (MAJOR.MINOR.PATCH). Each publish must bump this. |
type | enum | Yes | One of agent, claude-skill, cursor-rule, mcp-server, workflow, prompt-pack, custom-mode, tool-adapter, agent-template, automation. See Installable AI capabilities. |
description | string | Yes | One-line summary shown in search results. Keep it under ~120 characters. |
platforms | string[] | Yes | Target platforms: claude-code, claude-desktop, cursor, windsurf, openai-agents, gemini-cli, github-copilot, replit-agent, mcp. Drives compatibility badges and install targets. |
permissions | object | Yes | Declared permission scope. Each key is a boolean. Used to compute the risk level. See Permissions model. |
entry | string | No | Path to the entry file loaded by the tool. Defaults to instructions.md. |
license | enum | No | MIT, Apache-2.0, GPL-3.0, Proprietary, or Unknown. Defaults to Unknown. |
Notes on individual fields
platforms
List every platform your package genuinely supports — not aspirational ones. The marketplace renders a compatibility table from this field, and the CLI uses it to decide which install and export targets are valid. See Platform compatibility.
permissions
The permissions object is the heart of Nuclexa's trust model. Declare only what you truly need: fewer, lower-risk permissions earn a lower risk badge and more installs. Any key you omit defaults to false.
Next: the Permissions model explains each permission and how risk is computed.