AgentReverse

Extract tools, skills, and plugins from any source — GitHub repos, local configs, articles — and install only what you need. No cloning entire repositories for one function.

What's new in v2.0.0

Security scanning. Every install_capability call now runs through a security scanner before anything touches your config. Pattern-based (regex + AST), checks 6 categories: remote execution, data exfiltration, secret access, persistence, destructive operations, obfuscation. Critical findings block the install. Override with forceInstall if you know what you're doing.

Local introspection. local_scan inventories your entire Claude Code environment — skills, commands, hooks, MCP servers, keybindings, binary version. local_optimize finds dead skills, deprecated settings, duplicate scopes. Results are cached with SHA256 file hashes so repeat calls only rescan changed files.

Changelog awareness. changelog_check compares your Claude Code version against a local cache. When it detects a version bump, it fetches changelogs from npm + GitHub, matches against migration rules, and auto-applies safe changes. Breaking changes require your approval. No cost when the version hasn't changed.

30 MCP tools total (up from 26).

Installation

Quick start (Claude Code)

# Add MCP server
claude mcp add agent-reverse -- npx -y @shihwesley/agent-reverse agent-reverse-server

# Install slash commands (run from your project directory)
npx -y @shihwesley/agent-reverse setup

Restart Claude Code. You get:

• 30 MCP tools for extraction, analysis, security scanning, and local introspection
• /agent-reverse — reverse engineer capabilities from repos, local configs, articles, or your own setup
• /agent-reverse-update — check for skill updates and Claude Code version changes

Upgrading

# Re-run setup for latest skills (MCP auto-updates via npx)
npx -y @shihwesley/agent-reverse@latest setup

Global install (optional)

npm install -g @shihwesley/agent-reverse

# Then use without npx:
claude mcp add agent-reverse -- agent-reverse-server
agent-reverse setup
agent-reverse --help

From source

git clone https://github.com/shihwesley/agent-reverse.git
cd agent-reverse
npm install && npm run build

# MCP server
claude mcp add agent-reverse -- node dist/server.js

# CLI
node dist/cli.js setup
node dist/cli.js --help

How it works

AgentReverse is two things:

1. MCP server — fetches repos, parses ASTs, manages the manifest, runs security scans. This is the engine.
2. Skills layer — lives inside your agent (Claude Code, Cursor, etc.), watches your workflow, and suggests capabilities when friction is high.

graph TD
    User([User]) --> Agent[Host Agent]
    Agent -->|Invokes| AR[AgentReverse MCP]
    AR -->|Reverse Engineer| Repo[(External Repo)]
    AR -->|Surgical Write| Local[Local Skills/Configs]
    AR -->|Record| Manifest[agent-reverse.json]

The manifest (agent-reverse.json) tracks every installed capability — source repo, pinned commit, dependencies. Think package-lock.json but for agent skills.

Usage

Analyze a repo

/agent-reverse analyze https://github.com/cool-user/search-plugin

Returns a categorized bill of materials: every skill, tool, and prompt in the repo.

Analyze your local setup

/agent-reverse analyze local

Scans your Claude Code environment for dead skills, deprecated settings, and optimization opportunities.

Extract and install

/agent-reverse install search-lite --target claude

Installs the capability, updates CLAUDE.md, and pins the commit in your manifest.

Sync across machines

/agent-reverse sync

Re-installs all capabilities from your manifest into a fresh environment.

Security scanning

Every install runs through 6 security checks before writing files:

| Category | What it catches | |----------|----------------| | Remote execution | Shell spawns, eval, child_process | | Data exfiltration | Outbound HTTP with file reads, DNS exfil | | Secret access | Reads of .env, credentials, API keys | | Persistence | Cron jobs, startup scripts, auto-run hooks | | Destructive operations | rm -rf, file overwrites, git force-push | | Obfuscation | Base64 payloads, hex encoding, string construction |

Critical findings block the install. Medium findings warn. Low findings inform. Use forceInstall to override if you've reviewed the code.

Agent portability

Supported targets:

| Agent | Config path | Status | |-------|------------|--------| | Claude Code | .claude/skills/ | Supported | | Cursor | .cursor/rules/ | Supported | | Antigravity | skills/ | Supported | | Custom adapters | — | Planned |

Backup and restore

Export your setup for portability across machines or agent platforms.

What gets backed up

| Item | Path | |------|------| | Manifest | agent-reverse.json | | Skills | .claude/skills/*.md | | Commands | .claude/commands/*.md | | Rules | .cursor/rules/*.mdc | | CLAUDE.md | CLAUDE.md | | Caches | workflow-cache.json, known-repos.json |

Create a backup

# Local file (default: agent-reverse-backup-<date>.json)
agent-reverse backup

# Custom filename
agent-reverse backup --output my-backup.json

# Private GitHub Gist
agent-reverse backup --gist

# Public Gist
agent-reverse backup --gist --gist-public

# GitHub repo
agent-reverse backup --repo myuser/agent-backups

Restore from backup

# From local file
agent-reverse restore backup.json

# From Gist
agent-reverse restore https://gist.github.com/user/abc123

# From GitHub repo
agent-reverse restore https://github.com/user/backups/blob/main/backups/2026-01-26.json

Cross-agent restore

# Restore Claude skills to Cursor format
agent-reverse restore backup.json --target cursor

# Dry run
agent-reverse restore backup.json --target cursor --dry-run

Restore flags

| Flag | Description | |------|-------------| | --target <agent> | Convert to: claude-code, cursor, antigravity | | --merge | Merge with existing setup instead of replacing | | --force | Overwrite existing files | | --dry-run | Preview changes without writing |

MCP tools reference

| Tool family | Tools | Purpose | |-------------|-------|---------| | repo_* | repo_fetch, repo_analyze, repo_cleanup | Fetch and analyze GitHub repos | | manifest_* | manifest_add, manifest_remove, manifest_get, manifest_list, manifest_audit, manifest_sync, manifest_check_updates | Manage the capability manifest | | install_capability | — | Install with security gate | | conflict_* | conflict_check, conflict_resolve | Detect and resolve skill conflicts | | deps_* | deps_check, deps_resolve | Dependency analysis | | observer_* | observer_log, observer_get_patterns, observer_clear | Workflow pattern observation | | suggester_* | suggester_check, suggester_add_repo, suggester_list_repos | Capability suggestions | | web_* | web_fetch, web_interpret | Extract skills from articles and docs | | backup_* | backup_create, backup_restore, backup_list | Backup and restore | | security_scan | — | Pattern-based security analysis | | local_* | local_scan, local_optimize | Environment introspection | | changelog_check | — | Version-aware auto-migration |

Documentation

• PRD — vision and roadmap
• Tech spec — architecture and JSON schemas
• Test cases — real-world usage scenarios

Contributing

PRs for new agent adapters or language parsers are welcome. See the tech spec for architecture details.

License

MIT