MCP server that scans code for exposed secrets — API keys, tokens, private keys and high-entropy strings — with placeholder-aware allowlisting and redacted reports.
Strong tool definitions with clear, detailed descriptions (avg 180 chars) and well-structured schemas. All 6 tools have explicit input schemas with typed parameters and descriptions. Naming follows verb_noun convention (scan_*, list_*). Security-focused design with read-only operations and redaction of sensitive values. Main gaps: no error handling guidance, no output schema documentation, no parameter constraints (enums/ranges), and missing tool annotations (readOnlyHint). Composition is clean, each tool has one responsibility. Descriptions are LLM-optimized but could be more concise.
List the active secret detectors and allowlist rules.
Recursively scan a directory tree for exposed secrets. Automatically skips .git, node_modules, .venv/venv, __pycache__, dist, build, minified JS bundles, lockfiles, binaries, files over 5 MB, and simple patterns from the root .gitignore (best-effort: no negations, no ** globs).
Scan a single file for exposed secrets. Binary files (null-byte heuristic) and files larger than 5 MB are skipped and reported as such in the summary.
Scan the lines added by the most recent commits (git log -p). Each finding is tagged with the short hash of the commit that introduced it. A secret that was later deleted is still reported — history retains it, so the credential must be rotated regardless.
Scan ONLY the lines currently staged for commit (git diff --cached). This is the pre-commit checkpoint: it inspects exactly the content the next commit would publish, and reports the file and post-commit line number of every added secret.
Output schemas not documented in tool descriptions. LLMs cannot plan downstream operations or extract fields without knowing the response structure (e.g., findings array fields, summary format).
No error handling guidance. Descriptions do not explain what happens on failure (e.g., file not found, invalid git repo, permission denied) or how the LLM should recover.
Missing tool annotations. No readOnlyHint declared despite all tools being read-only operations. Agents cannot distinguish safe tools from destructive ones without explicit hints.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 61 | 2026-07-28+ | v2 |
Scan a snippet of text or code for exposed secrets. Detects API keys (GitHub, OpenAI, Anthropic, NVIDIA, AWS, Stripe, Google, Twilio), Slack/Discord tokens and webhooks, JWTs, private key blocks, credentialed connection strings, generic password/secret/token assignments, and high-entropy strings — while skipping obvious placeholders (allowlist).
Parameter constraints missing. max_files (scan_directory) and max_commits (scan_git_history) lack min/max bounds. LLMs could pass absurd values (0, negative, millions) causing timeouts or errors.
Descriptions exceed 200 chars for most tools (avg ~180), approaching token waste threshold. Trim to 100-150 chars while preserving key details (what, when, what it returns).