🛡️ Database Sentinel

Security audits for your database backend. Ask "audit my database" and get a scored report with exact fix code.

Works as a Claude Skill (Supabase, MongoDB), a read-only MCP server for any AI client, or a CLI agent that uses your own model (both Supabase).

→ MCP server setup and usage guide

→ Don't want to run it yourself? Locksoup is the hosted version: a Supabase security audit with scheduled re-checks and alerts.

Changelog

2026-10-01 (v1.0.1)

• Fixed: an UPDATE policy without WITH CHECK is no longer reported as letting users reassign row ownership (Postgres reuses USING as the check). Q8's label and the agent prompt said otherwise, which produced false mass-assignment findings on real projects.

• Benchmark: new dev case 027 from that project. Test F1 unchanged (0.836 vs 0.838, same day). Runs.

2026-09-30 (docs)

• MCP flow benchmarked: a model driving the real server over stdio scored test F1 0.865, precision 0.905, critical recall 1.0 (results). Harness: evals/mcp_client.py, eval system m.

v1.0.0 (2026-09-30): MCP server release

• MCP server v1.0: four read-only Supabase tools, an audit prompt and the catalog resources. Benchmarked end to end: a model driving the server over stdio scored test F1 0.865, with critical recall 1.0.

• Verified on hosted Supabase: session pooler, read-only role with BYPASSRLS, all 20 queries.

• Fewer false alarms:

  • Q1 now checks API grants: RLS off with anon/authenticated revoked isn't reported as exposed.

  • The audit prompt knows SECURITY INVOKER functions only have the caller's rights.

  • Found on a real project; new dev case 026.

• Pin a version: git+https://github.com/Farenhytee/database-sentinel@v1.0.0 in any uvx/pipx command. The Claude Code plugin is pinned automatically.

• Standard audit is now the CLI default: one prompt plus anon-probe verification, ~$0.0007 and ~15s per audit. --deep runs the tool-using agent, which scored the same on our bench at ~5× the cost.

v0.2.1 (2026-09-30)

• Test F1: agent 0.782 → 0.831, single-prompt 0.766 → 0.849. No crashes or timeouts (results).

• Fixed:

  • A bad tool argument no longer ends the audit; the error goes back to the model.

  • OpenRouter calls route to the fastest provider (slow ones caused timeouts).

  • The agent's final findings step no longer lists candidates it had dismissed.

v0.2.0 (2026-09-29)

• Published test results: agent F1 0.782, single-prompt 0.766, rules 0.559 on a blind, locked 10-case test split, 3 runs each (results).

• Bench: 25 labeled cases (15 dev, 10 blind test).

• Agent:

  • Verify step: probes as anon to confirm exploitable findings.

  • --fix: prints fix SQL for findings you pick. It's never executed.

  • Cost reporting: tokens, $ and tool calls for each audit.

• CI: free rules-only regression gate.

2026-09-29

• One-command install: Claude Code plugin (skill + MCP server, prompts for your connection string) and an Add to Cursor button.

• Install from GitHub: one command for the MCP server (uvx --from git+… sentinel-mcp) and the agent (pipx install "database-sentinel[agent] @ git+…").

• sentinel-mcp --role-sql prints the read-only role setup SQL.

• Clearer errors: the MCP client now sees why a call failed, and the CLI prints a single-line error.

2026-09-28

• New: MCP server for Supabase. Four read-only tools, an audit prompt and the pattern catalog as resources. Your own LLM does the analysis, so it's free.

• New: lite agent (sentinel-audit). A LangGraph pipeline that works with any OpenAI-compatible model: OpenRouter, OpenAI, or local Ollama.

• New: benchmark + evals. 10 labeled Supabase cases, with precision/recall/F1 compared across a rules baseline, a single-prompt baseline and the agent.

• Fixed: six Supabase audit queries. Q8 (UPDATE without WITH CHECK) never matched anything, Q16 and Q19 missed rows for least-privilege roles, and Q6, Q13 and Q14 had wrong conditions.

Related MCP server: schema-guard-mcp

Quick start

Claude Code installs the audit skill and the MCP server in one go (needs uv):

claude plugin marketplace add Farenhytee/database-sentinel
claude plugin install database-sentinel@database-sentinel

Supabase users: create the read-only login (step 1), then run /plugin configure database-sentinel@database-sentinel in Claude Code. Then ask: Audit my database.

Cursor  

Other MCP clients, and full setup: MCP guide

Skill only (Claude Code picks it up automatically):

git clone https://github.com/Farenhytee/database-sentinel.git ~/.claude/skills/database-sentinel

CLI agent (Supabase, bring your own model):

pipx install "database-sentinel[agent] @ git+https://github.com/Farenhytee/database-sentinel"
export SENTINEL_BASE_URL=http://localhost:11434/v1 SENTINEL_MODEL=<model>   # Ollama, or any OpenAI-compatible API + SENTINEL_API_KEY
sentinel-audit --dsn "postgresql://sentinel_auditor:<password>@<host>:5432/postgres" \
  --rest-url https://<ref>.supabase.co --anon-key <anon key> --repo ./my-app

The default audit is one prompt plus anon-probe verification. --deep runs a tool-using agent instead: same accuracy on our bench, ~5× the cost. --fix prints fix SQL for the findings you pick; it's never executed.

What it catches

Full tables are in Appendix A.

Safety

• Read-only by default. Only system catalogs are read. The MCP server and agent can't write, drop or delete anything.

• Your rows are never read. The audit role has no access to table data. Repo scans report file and line, never secret values.

• Write and network probes are opt-in (Skill only). Details are in Appendix D.

Status

License

MIT. Exception: database_sentinel/agent/, database_sentinel/mcp_server/ and database_sentinel/kb/ are AGPL-3.0 (see the LICENSE in each).

Appendix

A. Pattern highlights

Supabase

MongoDB

The MongoBleed probe (mongobleed-probe.md) is a single-packet, read-only detector. It was verified against mongo:7.0.20 (vulnerable) and 7.0.28 (patched), and runs only after two opt-in confirmations.

B. How the Skill audits

1. Detect backends. 2. Scan code for exposed credentials. 3. Introspect schema, policies and config. 4. Match against the anti-pattern catalogs. 5. Probe safely (opt-in). 6. Score and report. 7. Generate fix code.

Only SKILL.md (~2K tokens) and core/* load up front; each backend's files load only when that backend is detected.

C. Example output

╔════════════════════════════════════════════════════════╗
║                  SENTINEL SECURITY AUDIT               ║
║  Backends: supabase        Score: 35/100 🔴            ║
╚════════════════════════════════════════════════════════╝

🔴 CRITICAL — public.users: RLS Disabled                  [RLS_DISABLED]
  Risk:   Anyone on the internet can read your entire users table.
  Attack: Copy the anon key from DevTools → curl the API → dump all rows.
  Source: CVE-2025-48757 / Splinter 0013_rls_disabled_in_public
  Fix:    ALTER TABLE public.users ENABLE ROW LEVEL SECURITY;
          CREATE POLICY "users_select_own" ON public.users FOR SELECT
            TO authenticated USING ((SELECT auth.uid()) = id);

D. Safety details

• Supabase write probes: Prefer: tx=rollback, so no data is modified.

• MongoDB write probes: session + abortTransaction() on replica sets; canary insert + delete on standalone (opt-in).

• MongoBleed network probe: double opt-in; some monitoring tools alert on the 42-byte probe packet.

• Auth probes use .invalid email domains (RFC 6761).

• Credentials are held in memory for the audit only, and redacted in reports.

• MCP server: five independent read-only layers (details).

E. Benchmark and evals

bench/cases/ holds labeled Supabase schemas applied to a local Supabase. evals/ scores three systems on precision, recall and F1: R0 (rules, no LLM), B0 (single prompt) and A (agent). The test split is frozen and never tuned on.

Test split (10 blind, locked cases), deepseek-v4-flash, 3 runs each:

v0.2.1 fixes the crashes and timeouts seen in the v0.2.0 test run. The fixes were diagnosed on dev only, but they came from test failures, so v0.2.0 is the blind number (v0.2.1 details). A and B0 are within noise of each other. The agent is more precise, gives fewer false alarms on clean projects and verifies findings as anon, at ~5× the cost.

pip install -e ".[dev]" && supabase start && python -m evals.run --split dev

F. Continuous monitoring (GitHub Actions)

Templates: github-action-supabase.yml and github-action-mongodb.yml. They run on migration, rule or IaC changes, weekly, or manually, then comment on the PR and fail the build on critical findings. Ask Claude: "Set up continuous security monitoring for this project."

G. Repo layout

SKILL.md, core/, backends/{supabase,mongodb}/, references/   Claude Skill (MIT)
database_sentinel/mcp_server/                                MCP server (AGPL)
database_sentinel/agent/                                     lite agent (AGPL)
bench/, evals/, tests/, supabase/                            benchmark, eval harness, local Supabase
docs/mcp.md                                                  MCP setup guide
compat/supabase-sentinel/                                    old skill name shim

H. Research sources

CVE-2025-48757 (170+ Lovable apps), Escape.tech (2,000+ vulns in 5,600 vibe-coded apps), Veracode (45% of AI code has OWASP Top 10 issues), CMU SusVibes, ModernPentest (20.1M rows, 107 YC startups), Supabase Splinter lints, CVE-2025-14847 MongoBleed, Mongoose CVE-2024-53900 / CVE-2025-23061, CIS MongoDB 7 Benchmark. Full list: vibe-coding-context.md, cve-feed.md.

I. Contributing

Most valuable: new anti-patterns with evidence (CVE, breach report, or lint), better fix templates, false-positive/negative reports from live use, and new backends following backends/supabase/. Fork, branch, and open a PR with the pattern and its source. Contributions to the AGPL directories need a CLA.

J. Naming history

Supabase Sentinel (v1) → Database Sentinel (v3, multi-backend). The supabase-sentinel name still works through compat/supabase-sentinel/.