Docs/MCP Server

MCP Server

reel runs as a Model Context Protocol server, exposing its container-extraction tools to Claude Code, Cursor, Continue, and other MCP-aware AI clients. Local-only, stdio transport.


Install

Install the CLI, then add reel to your MCP client's config.

# macOS
brew install getreeldev/tap/reel
# Verify
reel start mcp --help

Claude Desktop / Claude Code

Add reel to ~/.claude/mcp_servers.json (Claude Desktop) or your project's .mcp.json (Claude Code):

{
  "mcpServers": {
    "reel": {
      "command": "reel",
      "args": ["start", "mcp"]
    }
  }
}

Cursor / Continue

Same JSON shape — both read the standard MCP server config. Restart the client after editing.

Seven tools

All artifact tools accept destination{ inline | local | s3 }. inline is capped at 1 MB so an oversize SBOM won't blow the AI's context. Discovery tools accept summary=true to return counts only.

ToolGroupWhat it does
whoamiIdentityMode, version, available tools, runtime.
healthIdentityLiveness probe for the running server.
list_workloadsDiscoveryLocal runtime containers. Supports summary=true.
list_imagesDiscoveryLocal images. Supports summary=true.
sbomExtractionCycloneDX SBOM. Scanners: vuln, secret, license, config, vex.
cbomExtractionCrypto Bill of Materials — algorithms, certs, keys.
malwareExtractionClamAV scan of the container or image rootfs.

Vendor VEX in the AI loop

Pass scanners: ["vuln", "vex"] on the sbom tool and the AI gets Trivy's findings already annotated with vendor verdicts (not_affected / fixed / affected) — pulled live from vex.getreel.dev. The model reasons about what's actually exploitable instead of the raw CVE list.

Coming next

Agent-mode MCP — an in-process listener on reel start server exposing the full agent surface (CRIU checkpoints, layer/frame/memory capture, file inventory) to AI clients over HTTP/SSE. Same tool registry, wired to the in-cluster extraction path.