Supported configuration

ModelFence evaluates explicit, implemented formats. Keep configuration in its documented location so discovery can associate it with the correct source.

MCP configuration

The analyzer recognizes .mcp.json, .vscode/mcp.json, .cursor/mcp.json, .gemini/settings.json and claude_desktop_config.json. Supported server definitions may launch a local command or name a remote endpoint.

A project-level .mcp.json might contain:

{
  "mcpServers": {
    "project-files": {
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-filesystem@2025.8.21", "/workspace/project"]
    }
  }
}

This is a configuration example, not a recommendation to install that package version. Choose and verify a version approved by your organization. Pinning improves reproducibility; it does not prove package safety. Restrict filesystem paths to the project rather than a home directory or the host root.

Checks include unpinned packages, literal credentials, non-loopback unencrypted HTTP, broad filesystem scopes and privileged Docker launch options. Use the rule catalog to understand the exact conditions.

Coding agent permissions

Claude Code project settings in .claude/settings.json are inspected for bypassed permission prompts, unrestricted shell access and automatic approval of project MCP servers. The source configuration may be broader than the permissions actually available to a deployed process; findings remain labeled Configured unless separately verified.

GitHub Actions AI agents

The workflow analyzer identifies supported Claude Code, Codex and Gemini CLI actions in .github/workflows/. It evaluates token permissions, untrusted triggers, action pinning and sandbox-related settings covered by its rules.

What is outside this analysis

Arbitrary LangChain, Semantic Kernel and other framework code is not analyzed as a complete permission model. Cloud agent services, model weights and prompt-injection testing are not implemented capabilities. Unsupported formats remain unsupported in Coverage.

After editing a configuration, sync its integration or rerun the relevant passive assessment. Review the new scan’s evidence instead of assuming the edit was deployed or took effect in production.