agentver

Team-grade skill management for AI coding assistants.
Version, audit, and deploy skills to 43+ assistants — with access control, security scanning, and a review workflow built in.

Website · Platform · Docs · Quick Start

What is Agentver?

Agentver is an open-source platform for managing agent skills across teams. Write a skill once and deploy it to every assistant your team uses — Claude Code, Cursor, Windsurf, GitHub Copilot, Gemini CLI, and 38 more — with consistent versioning, access control, and security scanning built in.

It is designed for teams: from engineers who want reproducible agent configurations, to technical leads who need to control what skills are deployed and by whom, to non-technical contributors who can create and share skills through the web dashboard without touching a config file.

Key features

• Team management — role-based access control (owner, admin, member, viewer) with org and sub-team structure, invitation workflows, and a full audit log
• 43+ agent support — one skill, every assistant; auto-detects installed agents and places files in the right location
• Git-native versioning — semantic versions, commit SHAs, lockfiles, and manifest files committed to your repo so teams stay in sync
• Security scanning — 75 built-in detection rules across 9 threat categories; runs locally with no external API calls; blocks installs on HIGH or CRITICAL findings
• Credential vault — AES-256-GCM encrypted secrets with team/org sharing, rotation policies, expiry, and access logs
• Bundles — group skills, MCP servers, scripts, and config into installable workflow collections
• MCP server catalogue — browse and connect MCP servers to your coding assistants, bundled with skills for one-command setup
• Change proposals — PR-style review workflow for proposing and approving improvements to shared skills
• Community discovery — search and browse published skills from the community
• Self-hosting — run the full platform on your own infrastructure with Docker

Quick start

CLI

# Install
npm install -g @agentver/cli
# or: bun install -g @agentver/cli

# Install a skill from a Git repository
agentver install github.com/owner/repo

# Scan your project for existing agent configs
agentver scan

# Audit installed skills for security issues
agentver audit

# Check status of installed skills
agentver status

# List installed skills
agentver list

See the CLI documentation for the full command reference.

Self-hosting (Docker)

git clone https://github.com/agentver/agentver.git
cd agentver/docker
./setup.sh

Your dashboard will be available at http://localhost:3000 within minutes. See the docker/ directory for configuration options.

Cloud

Use the hosted version at app.agentver.com — no setup required.

Security scanning

Every skill install is scanned automatically before any file is written to disk. The built-in scanner runs entirely locally — no external API calls, no data leaving your machine.

75 detection rules across 9 categories:

Verdicts: PASS (proceed) · WARN (confirm before install) · BLOCK (halted — fix before retrying)

The scanner uses deterministic rule-based pattern matching, not AI analysis — it cannot be manipulated by a skill that embeds a prompt injection string to trick an AI reviewer.

You can configure trusted source patterns to skip scanning for your own verified repositories, and adjust the block severity threshold in .agentver/config.yaml.

Repository structure

apps/
  dashboard/          # Next.js 16 — web dashboard, skill browser, import gateway
  website/            # Marketing site
  desktop/            # Tauri 2.x desktop app
packages/
  cli/                # @agentver/cli — CLI tool (MIT)
  shared/             # @agentver/shared — schemas, types, validation (MIT)
  agent-definitions/  # @agentver/agent-definitions — 43+ agent configs (MIT)
  mcp-server/         # @agentver/mcp-server — MCP server integration (MIT)
  github-action/      # @agentver/github-action — CI/CD action (MIT)
  database/           # Prisma 7 + PostgreSQL schema
  ui/                 # Radix/shadcn components
  ui-utils/           # Shared UI utilities
  typescript-config/  # Shared TypeScript configs
docker/               # Self-hosting Docker Compose setup
e2e/                  # Playwright end-to-end tests

Tech stack

FAQ

Is this just npm for prompts?

No. npm is a package registry for code. Agentver is a team management platform for agent skills. The differences that matter for teams: role-based access control so you can control who publishes and installs skills; git-native versioning with lockfiles so your whole team installs the same version; a security scanner that blocks malicious skills before they touch disk; an encrypted credential vault for API keys used by MCP servers; and a change proposal workflow so skills can be reviewed before they're deployed. A prompt file in a git repo solves a different problem.

Does this overlap with GitHub Actions?

No. GitHub Actions is a CI/CD pipeline runner — it executes scripts when you push code. Agentver manages what your AI assistants know and can do, which is a separate concern. The two work well together: use the Agentver GitHub Action to keep skills in sync as part of your CI pipeline.

How do you keep up when agents like Claude, Cursor, and Windsurf update their formats?

Each agent's configuration paths and file formats are defined in a data layer (packages/agent-definitions). When an agent changes its config format or adds a new path, we update one entry — the skill content itself doesn't change. Skills declare which agents they target; the platform handles translation and placement. Adding support for a brand new agent is a data change, not an architectural one.

How is this different from LangChain Hub or Smithery?

LangChain Hub is a prompt store for LangChain-specific chains. Smithery is an MCP server marketplace. Neither manages cross-agent skill distribution, team permissions, security scanning, credential storage, or version lockfiles. Agentver is closest to a private package registry combined with a team configuration manager — the goal is consistency and control across an entire engineering team, not just individual developer tooling.

Is the security scanner actually reliable?

The scanner has 75 deterministic rules covering dangerous commands, data exfiltration, credential exposure, code obfuscation, Unicode-based evasion techniques, prompt injection, and remote code execution. It is not AI-based — it uses rule-based pattern matching, which means it cannot be fooled by a prompt injection payload designed to manipulate an AI reviewer. No static scanner is exhaustive, and we say that plainly: it is a defence-in-depth layer, not a replacement for code review. The rules, verdicts, and configuration are all open source — you can read exactly what it checks.

Can I self-host?

Yes. The full platform runs on Docker Compose. See the docker/ directory. The CLI packages are MIT-licensed and work completely standalone without any platform connection.

Contributing

See CONTRIBUTING.md for development setup and guidelines.

Licence

• Platform (apps, database, UI): AGPLv3
• CLI packages (cli, shared, agent-definitions, mcp-server, github-action): MIT