taniwhaai/arai
Policy enforcement for AI coding agents derived from existing instruction files (CLAUDE.md, .cursorrules, .windsurfrules, .github/copilot-instructions.md) — no separate YAML to maintain. Rules with prohibitive predicates (never, forbids, must_not) emit permissionDecision: deny to block tool calls in Claude Code; advisory rules inject context. PostToolUse is correlated with PreToolUse to produce per-rule obeyed/ignored compliance verdicts in a local JSONL audit log. MCP tools — arai_add_guard (register rules mid-session), arai_list_guards, arai_recent_decisions — work in any MCP client (Claude Desktop, Cursor, Windsurf, Cline). No network on the hook hot path; opt-out anonymous telemetry.
Ads
Overview
An instruction file is advice — the model can read CLAUDE.md and still force-push anyway. Arai turns instruction files (CLAUDE.md, AGENTS.md, .cursorrules, and others) into enforcement via native hooks: rules derived from prohibitive language block the tool call outright, advisory rules inject the relevant constraint at the point it applies, and a tamper-evident audit log records, per rule, whether the model actually complied.
Installation
docker run --rm -i -v "$(pwd)/.taniwha/arai:/home/arai/.taniwha/arai" arai
Configuration
{
"mcpServers": {
"arai": {
"command": "docker",
"args": [
"run",
"--rm",
"-i",
"-v",
"$(pwd)/.taniwha/arai:/home/arai/.taniwha/arai",
"arai"
]
}
}
}Paste into ~/.cursor/mcp.json or .cursor/mcp.json
Tags
- awesome-list
Related MCP servers
Cryptography
ucsandman/DashClaw
Fail-closed approval layer for unattended agent runs: `guard` evaluates each declared action against org policy before it executes (allow/warn/block/require-approval with one-click human approval), records every decision to a causal ledger, and adds plan preflight, scoped delegation grants, and containment verdicts. `npx -y @dashclaw/mcp-server`
Cryptography
luiacuaniello/perspectivegraph
Attack-path engine for cloud and Kubernetes. Eight read-only tools let an agent list the reachable routes from internet exposure to sensitive assets, explain each hop and the evidence behind its probability, find choke points, and simulate cutting a relationship before recommending the fix. Connects to a running PerspectiveGraph: `perspectivegraph mcp --api http://localhost:8080`.
Cryptography
Gowthaman90/mcp-bastion
Reliability + security proxy that sits in front of MCP servers: TOFU tool-definition pinning against rug pulls, tool-poisoning and cross-server-exfiltration detection, argument/command-injection blocking, inline secret redaction, MCP 2026-07-28 header/body validation (`-32020`) and cache-policy clamping, and a compliance-mapped audit trail (NIST AI RMF / OWASP). Coverage measured on an open benchmark with a held-out corpus. `npx -y mcp-bastion`
Cryptography
elberacasa/umbra
Trust score and guardrails for AI-generated code: static security rules, Docker-verified build/boot checks, and claim receipts that catch agents lying about tests. Tools: `scan_repo`, `guard_content`, `get_score`. Run with `npx --yes -p @elberacasa/umbra umbra-mcp`.
Cryptography
M2M-Sentinel/m2m-sentinel-sdk
Deterministic EVM bytecode capability intelligence, EIP-1967 proxy resolution, gas recommendations, and preflight safety intelligence for autonomous agents on Base (`Chain ID: 8453`). Supports stdio and hosted SSE. `npx -y m2m-sentinel-sdk`
Cryptography
Drumworks/ssid-mcp
MAC-address (OUI) vendor lookup and router default-login directory for AI agents. Identify a device manufacturer from its MAC address, detect randomized/private (locally-administered) addresses instead of reporting "unknown", or fetch a router's default login IP and admin credentials — every router field cited to the manufacturer's own documentation. Free tier, no signup. `npx -y ssid-mcp`
Cryptography
aggrete/aggrete
Policy proxy that governs what AI assistants can reach and do: refuses forbidden tool calls before the upstream is contacted, with per-user memory and a tamper-evident audit. Stops prompt-injection exfiltration and forbidden data combinations across Slack, Drive, GitHub and more. `pip install aggrete`.
Cryptography
SaravananJaichandar/etch-mcp
Signed audit chain for AI agent decisions. Every event signed, Merkle-chained per project, anchored to public transparency logs, and offline-verifiable against a pinned public key without dependency on our infrastructure. Post-hoc evidence primitive, complementary to runtime enforcement (not a substitute). Zero-signup try-it-now: `curl -X POST https://etch.systems/v1/your-project` returns a bearer token you can drop into any MCP client.