openaca 0.1.0b3

OpenACA: Agent Composition Analysis — overlays and scanner for AI agent security.

OpenACA

Agent Composition Analysis (ACA) — OSV-compatible agent-context overlays and a reference scanner for AI agent infrastructure: plugins, MCP servers, skills, agent frameworks, model proxies, and runtime components.

OpenACA is the AI-agent analogue of Software Composition Analysis (SCA): it identifies the versioned plugins, MCP servers, skills, and framework components that make up an AI agent, and matches them against known security records (CVE/GHSA/OSV + agent-context overlays maintained in this corpus).

Beta status

This is the 0.1.0b3 closed-beta pre-release. The scanner and overlay corpus are usable; expect rough edges. If you're a beta tester, start with the beta-tester guide — it covers install, first scan, what feedback I'm looking for, and how to report.

Install

Requires Python 3.11+.

Recommended — uv tool (install uv if you don't have it; uv also provisions Python for you so a 3.11+ runtime isn't a prerequisite you need to satisfy separately):

uv tool install openaca
openaca --version

Alternative — pip (if you already have a Python 3.11+ workflow):

pip install openaca

Both commands auto-pick the latest pre-release while OpenACA has no stable version yet. Current latest is 0.1.0b3; check with openaca --version.

Pin to a specific build if you need to reproduce a bug report: uv tool install openaca==0.1.0b3 or pip install openaca==0.1.0b3.

Try it in 30 seconds

mkdir openaca-demo && cd openaca-demo
cat > mcp.json <<'EOF'
{
  "mcpServers": {
    "git": {
      "command": "npx",
      "args": ["@cyanheads/git-mcp-server@1.1.0"]
    }
  }
}
EOF
openaca scan repo --target .

Expected output:

Found 1 vulnerability in 1 package.

@cyanheads/git-mcp-server 1.1.0
  location: mcp.json
  fix:      upgrade to >=2.1.5

  HIGH  GHSA-3q26-f695-pp76  fixed in 2.1.5  @cyanheads/git-mcp-server vulnerable to command injection in several tools  [osv.dev]

Scanned 1 manifest, 1 component. Sources: osv.dev.

For more scenarios (clean scan, configuration-hygiene checks via --include-posture), clone openaca-demo and try each of its fixtures.

Two scan modes

Mode	Question	Where it runs
openaca scan repo --target <path>	What agent components are declared in this repository?	CI gate, PR check
openaca scan endpoint	What agent components are installed on this machine right now?	Developer laptop, CI runner

Both modes emit text (default), JSON (--format json), or SARIF 2.1 (--sarif <path>). Use -v for per-finding context.

Configuration-hygiene posture rules (mutable install references, insecure transport, missing remote auth) are opt-in via --include-posture. They run separately from vulnerability findings and never fail CI by default.

V0 scope

• Endpoint mode scans Claude Code only (~/.claude or $CLAUDE_CONFIG_DIR). Other agent hosts aren't endpoint-supported yet.
• Repo mode parses Claude Code's declared manifests (.claude-plugin/plugin.json, .claude/settings.json) plus the host-agnostic mcp.json that most MCP-aware hosts use.
• Declared manifests only. SDK-inline definitions (query({ mcpServers: { ... } })), tools registered programmatically, and source-code parsing are V1 scope.
• Overlay corpus focuses on malicious-package records for MCP servers, agent-framework packages, and AI infrastructure. Pair with a general-purpose SCA scanner for your full dependency tree.

Links

• Beta-tester guide: https://github.com/open-agent-security/openaca-demo/blob/main/BETA-TESTER-GUIDE.md
• Sandbox fixtures: https://github.com/open-agent-security/openaca-demo
• Feedback: DM the maintainer ([e-mail]). The openaca source repo is private during the closed beta; GitHub issues open up when the repo flips public.

Apache-2.0.