samvallad33/vestige
Deterministic Operating System, Safety Kernel, and Causal Nervous System for AI Agents.
What it solves
AI coding agents often forget context from earlier sessions and can execute dangerous tool calls. Vestige gives them a persistent, local memory and a safety layer that blocks harmful actions before they run.
How it works
Vestige runs entirely on your CPU with no LLM, GPU, or cloud calls. It stores session histories and retrieves relevant memories on demand, using a method that outperforms plain BM25 across several benchmarks. Before each tool call, a set of 102 fixed rules checks for dangers like recursive deletes or credential leaks. Owner permissions expire after 15 minutes, so the agent must ask again. Every write is signed and hash-chained, producing an offline-verifiable audit log.
Who it’s for
Developers working with AI coding agents who need durable memory and protection against bad tool calls, and who prefer a fully local, CPU-only solution.
Highlights
- Persistent memory across sessions, allowing agents to trace failures back to the change that caused them.
- 102 fixed safety rules that block hazardous tool calls before they execute.
- One-time owner permits last 15 minutes, forcing re-approval.
- Signed and hash-chained writes for offline verification of every action.
- Benchmarked retrieval results beat plain BM25 on multiple datasets with zero embeddings and zero cloud spend.
- Runs on a plain CPU with 8 GB RAM, no LLM or GPU required.
Related
- Project
gendigitalinc/sageSage is a lightweight security layer that intercepts AI coding agents' tool calls (shell commands, URL fetches, file writes) and checks them against multiple threat-detection layers — cloud URL reputation, 300+ local YAML heuristics, ML-based prompt-injection detection, npm/PyPI supply-chain checks, plugin scanning, and Windows AMSI — before they execute. It supports Claude Code, Cursor, VS Code, OpenClaw, OpenCode, pi, and Oh My Pi, and is licensed under Apache 2.0 (source) with a separate Detection Rule License for its threat rules.
- Project
Justin0504/AegisAEGIS is a pre-execution firewall for AI agents: it intercepts every tool call, classifies it in real time, blocks or holds violations, and writes a tamper-evident audit trail — with one line of code and no changes to your agent. It includes a dashboard (Compliance Cockpit), cloud/local deployment options, and claims to cut attack success rates dramatically in benchmarks.
- Project
sangrokjung/claude-forgeClaude Forge is a plug‑in pack that turns the basic Claude Code terminal coding assistant into a production‑grade development environment. It adds 16 specialist AI agents, 35 slash‑commands, 33 saved “skills”, 22 safety hooks, 14 behavior rules and 4 external‑tool connections. Install with a one‑liner script (full install) or via Claude Code’s plugin marketplace (commands only). The pack automates planning, test‑first development, security reviews, documentation sync, and an adversarial verification loop that requires an independent agent to approve every change. Licensed MIT, actively maintained (v4.3.0, Sep 2026).
- Project
VibeTensor/attestixAttestix is an open-source toolkit that gives AI agents signed identities, verifiable credentials, and compliance evidence—helping developers meet AI governance requirements like the EU AI Act. It combines agent identity, provenance tracking, and trust scoring with MCP, REST, and CLI interfaces.
- Project
Juneo-Lab/decision-forgeDecision Forge is an open-source tool that compiles an organization's policy documents into verified decision training data for AI agents. It extracts rules, generates scenarios, labels them deterministically, and exports training datasets with provenance and leak-resistant splits. The core pipeline runs offline after human review, requiring no API keys. It supports multiple document formats and provides four export formats for model training.