Reins MCP Server
Audits Claude Code standing permission grants, scores each grant's prompt-injection blast radius, and generates deny-rule guardrails — local, read-only by default, zero network egress.
Baseline review at submission: an accuracy and overt-behavior check, not a full security audit, and it does not cover later changes to the repository. Learn more about our security review process.
Tools (7)
- reins_audit
Audit standing permission grants across Claude Code settings sources, classified into 12 risk categories with redacted evidence
- reins_explain
Deep-dive on a single grant: what it permits, why it was flagged, and remediation options
- reins_blast_radius
Rank grants by worst-case prompt-injection blast radius (credential access, exfiltration, destruction, persistence, code integrity, privilege escalation)
- reins_generate_policy
Generate permissions.deny rules, PreToolUse hook scripts, and a managed-settings baseline from findings (dry — never writes)
- reins_snapshot
Save the current grant set as a baseline for drift detection
- reins_drift
Diff live grants against the saved baseline; newly added grants are fully classified and scored
- reins_revoke
Remove selected allow rules by audit-issued rule id — confirmation-gated, timestamped backups, atomic writes, read-only sources refused
Every time a developer clicks “always allow,” Claude Code writes a standing permission into settings files that are almost never re-read. Those allowlists accumulate embedded credentials, destructive wildcards, credential-store reads, and unrestricted push/publish rights. Reins audits that surface before a prompt-injected agent session gets to use it.
What it does
Reins classifies every standing permission grant across Claude Code’s settings sources into 12 risk categories, using shell-aware tokenization (shlex) rather than regex-only matching so obfuscated or compound commands still classify correctly. Every non-safe finding gets a blast-radius score answering one question: if a prompt-injected agent fired this grant right now, what happens with no human in the loop? Findings convert directly into enforceable guardrails: permissions.deny rules, PreToolUse hook scripts, and an org-distributable managed-settings.json baseline. A snapshot/diff mechanism means a re-audit tells you exactly which grants appeared since your last known-good baseline.
How it works
The classification engine is pure stdlib Python with no I/O beyond the files it’s told to read or write, so it’s fully unit-testable independent of the MCP transport. Detectors are data: a table of regex patterns per risk category, so contributors add coverage without touching pipeline code. Blast-radius scoring is deterministic (severity × breadth × reach, no LLM in the scoring path) and every string a tool returns passes through a single redaction boundary before it leaves the process — known secret shapes and high-entropy tokens are replaced with a stable fingerprint so the same secret is recognizable across findings without the value ever appearing. Writes are confirm-gated: reins_revoke requires an explicit confirm: true, re-verifies each targeted rule is still present before touching anything, writes a timestamped backup, and performs an atomic os.replace. Removal is non-destructive — Claude Code simply re-prompts the next time the permission is needed. The server is stdio-only with zero network egress, enforced by an automated socket-blocking test across the full suite.