Offline, read-only auditing of MCP client configuration files with redacted findings
Runs on your machine from its package; the steps are below.
Not assessed yet.
Odel charges nothing per call. The publisher may charge for its own service.
Runs locally over stdio
This server isn't hosted — your MCP client launches it from a package registry. Use one of the commands below, or drop the config into your client (e.g. Claude Desktop).
Run
uvx mcp-latchpoint
MCP client config
{
"mcpServers": {
"latchpoint": {
"command": "uvx",
"args": [
"mcp-latchpoint"
]
}
}
}mcp-latchpoint audits MCP client configuration files without running the configured servers. It works offline, reads only local files you select, and produces text, JSON, or SARIF results.
This is an early defensive tool. Review findings in context before changing a working configuration.
The v0.1 rules cover:
npx and uvx packages without exact versionsEvery finding has a stable rule ID, severity, location, remediation, redacted evidence, and a confidence note when interpretation depends on the launched server.
Python 3.11 or newer is required.
python -m pip install mcp-latchpoint
To install from a local checkout:
python -m venv .venv
# Linux or macOS
. .venv/bin/activate
# Windows PowerShell
.venv\Scripts\Activate.ps1
python -m pip install .
For development:
python -m pip install -e ".[dev]"
pytest
ruff check .
mypy
The MCP server uses the official Python SDK v2 and the dependency is constrained to mcp>=2.3,<3.
Scan one or more explicit files:
mcp-latchpoint scan ~/.config/Code/User/mcp.json
mcp-latchpoint scan examples/risky.json --format json
mcp-latchpoint scan examples/risky.json --format sarif --fail-on high > results.sarif
Restrict every explicit path to an approved directory:
mcp-latchpoint scan ./configs --allowed-root ./configs
Directories are searched only for recognized MCP config filenames, up to 256 files. A file is limited to 1 MiB by default. Change these bounds with --max-files and --max-bytes.
Exit codes are 0 for a completed scan below the chosen threshold, 1 when a finding meets --fail-on, and 2 for input, containment, or parse errors. The default --fail-on none reports findings without failing a build.
List and explain rules:
mcp-latchpoint rules
mcp-latchpoint explain MCP004
Explicit files may use these structures:
| Client layout | Container | Accepted file syntax |
|---|---|---|
Claude Desktop, Cursor, portable .mcp.json, generic MCP JSON | top-level mcpServers object | JSON |
| Claude Code user/local settings | top-level mcpServers, plus projects.<path>.mcpServers in ~/.claude.json | JSON |
| VS Code | top-level servers object | JSONC for .vscode/mcp.json |
| Codex | [mcp_servers.<name>] tables | TOML |
Files ending in .jsonc are also parsed as JSONC. Other .json files remain strict JSON so malformed input is not silently accepted.
--discover checks only the following paths when they exist. It does not search the rest of the home directory.
~/.claude.json, ~/.cursor/mcp.json, ~/.codex/config.toml, $COPILOT_HOME/mcp-config.json with ~/.copilot/mcp-config.json as the fallback.mcp.json, .codex/config.toml, .cursor/mcp.json, .vscode/mcp.json%APPDATA%\Claude\claude_desktop_config.json, %APPDATA%\Code\User\mcp.json~/Library/Application Support/Claude/claude_desktop_config.json, ~/Library/Application Support/Code/User/mcp.json$XDG_CONFIG_HOME/Code/User/mcp.json, falling back to ~/.config/Code/User/mcp.jsonThe stdio server exposes two tools: list_rules and scan. It requires an allowed root at startup. Relative scan paths are resolved below that root; absolute paths, .. traversal, and symlinks cannot escape it.
mcp-latchpoint-server --root /absolute/path/to/reviewed-configs
mcp-latchpoint serve --root /absolute/path/to/reviewed-configs
Example client entry:
{
"mcpServers": {
"latchpoint": {
"command": "/absolute/path/to/mcp-latchpoint-server",
"args": ["--root", "/absolute/path/to/reviewed-configs"]
}
}
}
The server returns findings and scan metadata, never raw configuration content. Stdio is the only server transport exposed by the entry point, and stdout is reserved for MCP protocol messages.
The serve command is also the entry point advertised in the MCP Registry. Set --root to a directory you explicitly trust before connecting a client.
The Glama listing builds a container from the repository. In its Dockerfile configuration, use Python 3.13, these build steps and command arguments:
["uv sync --no-dev"]
["mcp-proxy", "--", "/app/.venv/bin/mcp-latchpoint-server", "--root", "/app/examples"]
The root is an existing directory with synthetic sample configurations. It lets Glama start and inspect the tools without giving the server access to a user's files. The command uses the executable inside the virtual environment created by uv sync. For this demo, the environment-variable schema can be {"type":"object","properties":{}} and placeholder parameters can be {}.
To scan your own configurations, run the server locally with --root pointing to a directory you explicitly trust. Glama's demo root is only for the sample files in examples/.
The scanner never executes commands, installs packages, resolves referenced environment variables, or connects to endpoints. It does not follow configuration includes or inspect an MCP server's code or runtime behavior. Argument-based rules are intentionally limited to recognizable patterns, so custom flags can be missed. A clean report is not proof that a server is safe.
Secret detection is designed to emit field names and <redacted> markers rather than values. If you find a leak or a path-containment problem, follow SECURITY.md and do not attach a real configuration to a public issue.
MIT. See LICENSE.