SieveAI-dev

Sieve — Security skill for Claude Code

Security community

A fully local LLM-traffic security proxy for crypto-native developers — redacts secrets outbound, blocks dangerous tool calls inbound (fail-closed).

How to install Sieve

This entry records only its repository, not the path inside it, so there is no exact command to give. Open SieveAI-dev/sieve and copy the folder into ~/.claude/skills/, or the file into ~/.claude/agents/.

What Sieve does

A fully local LLM-traffic security proxy for crypto-native developers — redacts secrets outbound, blocks dangerous tool calls inbound (fail-closed).

Alternatives in Security

  • Mcp-scan (Invariant Labs) — MCP security scanner with proxy mode for real-time scanning without infrastructure changes 1.9k ★
  • ThinkWatch — Self-hosted AI API and MCP gateway for organizations: SSO and RBAC, per-user identity for MCP tool calls, PII 817 ★
  • Agentseal — Security toolkit for AI agents 344 ★

README

Sieve

English | [中文](./README.zh-CN.md)

[![License: Apache-2.0](https://img.shields.io/badge/license-Apache--2.0-blue.svg)](./LICENSE) [![Platform: macOS](https://img.shields.io/badge/platform-macOS-lightgrey.svg)](#installation-phase-1-macos-only) [![Status: alpha](https://img.shields.io/badge/status-0.1.0--alpha-orange.svg)](#project-status)

**A fully local LLM-traffic security proxy — the last gate before irreversible actions.**

Sieve is a fully local LLM-traffic security proxy (a single Rust binary) that sits between your AI coding agent (Claude Code / OpenClaw / Hermes / Codex CLI) and the upstream model API (Anthropic / OpenAI / relays). It inspects traffic in both directions — redacting secrets on the way out, and blocking dangerous tool calls on the way in (fail-closed) — to force a moment of cognitive friction before irreversible actions (signing, transfers, deploys). Built for crypto-native developers.

All detection reasoning runs 100% on your machine. Sieve never uploads your prompts, responses, or API keys.


How it works

Your agent points its base URL at a loopback listener (`127.0.0.1`). Sieve forwards traffic to the real upstream over end-to-end TLS — it does **not** install a local CA or MITM your connection. On the way out it matches outbound rules and redacts secrets in place; on the way in it intercepts Critical tool calls and holds them for human confirmation (HIPS — Host Intrusion Prevention System).

flowchart LR
    A["AI agent
Claude Code / OpenClaw / Hermes / Codex CLI"] S["Sieve
(local · 127.0.0.1)"] U["Upstream
Anthropic / OpenAI / relay"] A -- "request" --> S S -- "outbound: redact secrets
(auto, no popup)" --> U U -- "response" --> S S -- "inbound: block Critical tool calls
fail-closed + HIPS confirmation" --> A S -.-> D["100% local detection
no cloud verifier"]

The only outbound traffic Sieve itself makes is to fetch signed rule updates — see [Privacy