# DROS VajraClaw (Hacker Edition)

> Free Docker governance gateway that enforces deterministic, sub-microsecond AI agent execution security via O(1) bitmap policies and cryptographic audit chains.

DROS VajraClaw (Hacker Edition) is a free, standalone Docker governance gateway built by Top-Celestial Company Ltd. for individual developers and AI researchers who need deterministic runtime security across multi-agent ecosystems. It enforces execution policy at the C-ABI/FFI boundary using compile-time resolved, cryptographically signed capability credentials rather than probabilistic LLM-based guardrails. The project is licensed under Apache 2.0 for individual use and is available on GitHub.

## What It Is

DROS VajraClaw Hacker Edition is a local Docker container (listening on `:8080`) that acts as a governance gateway between autonomous AI agents and the host operating system. Rather than relying on prompt engineering, LLM-as-a-judge, or semantic text matching, it moves security intelligence to compile-time: policies are declared in Markdown (`Vajra.md`) or YAML (`demo_policy.yaml`), compiled into deterministic `policy.bin` files signed with Ed25519, and enforced at runtime via O(1) bitmap lookups. The project positions itself as the "missing execution layer" for the Agentic Web — sitting at the FFI choke point where all agent-initiated OS system calls must pass through.

## How the Enforcement Architecture Works

The core enforcement model has three stages:

- **Compile-time:** Wildcards and capability inheritance are statically expanded; the resulting policy is hashed (SHA-256) and signed (Ed25519) into a `policy.bin` binary.
- **Runtime:** Every agent action that would cross the C-ABI boundary is checked against an in-memory O(1) bitmap. The project reports a core check latency of 484.8 ns and an end-to-end FFI overhead of 8–80 μs depending on platform.
- **Fail-closed:** Any capability not explicitly declared `ALLOW` is permanently blocked. Unauthorized calls trigger an OS-level Thread Panic (SIGKILL), not a soft rejection.

The audit trail is a SHA-256 Merkle hash-linked chain signed with Ed25519, designed to satisfy non-repudiation requirements the project associates with EU AI Act Article 12.

## Agent Ecosystem Integrations

The Hacker Edition ships with working starter templates for five major agent ecosystems:

- **Anthropic Claude Desktop & Claude Code** via MCP protocol (`/mcp` endpoint)
- **Cursor IDE / VS Code Agents** via `.cursorrules` terminal interception
- **OpenAI SDK & LangChain** via a 3-line Python `VajraClaw` wrapper
- **CrewAI & Microsoft AutoGen** with per-role W3C DID assignment and capability bitmap enforcement
- **DeepSeek Harness (DSH)** via a dual-mode plugin (`dsh-plugin-vajraclaw`)

The gateway exposes both REST (`/evaluate`) and MCP endpoints, making it compatible with any agent framework that can make HTTP calls.

## Policy-as-Code Developer Experience

Policies are written in either plain Markdown (`Vajra.md`) or structured YAML. The project provides an AI prompt template so developers can generate a policy by pasting a description of their agent's role into ChatGPT, Claude, or Cursor. Hot reloading is supported: mounting a new policy file takes effect in under 1 microsecond without restarting the container. The project recommends locking policy files to read-only (`chmod 444`) after configuration to prevent a compromised agent from rewriting its own rules.

## Technical Foundations and Benchmarks

The project references three Zenodo-archived technical papers covering its six-trust-boundary architecture, 4-layer defense-in-depth model, and runtime attribution framework. The project's own published benchmark reports 160,611 verified requests over a 24-hour soak test with a 26.1 μs decision latency and zero race conditions under concurrent RCU stress. The Hacker Edition supports up to 5 concurrent agents on a single host; out-of-the-box community mode supports 2 concurrent agents without a license key.

## Current Status

The repository was created in May 2026 and last pushed in August 2026, indicating active early development. The project is at v1.0.0 and is available as a pre-built image on GitHub Container Registry (GHCR). The broader DROS platform spans multiple language-specific core repositories (Rust, Go, C++, Python, Java, TypeScript) and is protected under U.S. Provisional Patent Application No. 64/111,973 (Patent Pending). Enterprise and startup commercial tiers exist separately from the free Hacker Edition.

## Features
- O(1) bitmap policy enforcement at C-ABI/FFI boundary
- Sub-microsecond in-band fusing (<1μs)
- Ed25519 cryptographic policy signing
- SHA-256 Merkle hash-linked audit chain
- W3C DID (did:key) agent identity binding
- RFC-010 passport minting and cross-agent verification
- Strict fail-closed enforcement (Thread Panic on breach)
- Policy-as-Code via Vajra DSL (Markdown or YAML)
- Hot policy reloading without container restart
- Up to 5 concurrent agents (Hacker Edition)
- MCP and REST API endpoints
- Docker-based deployment
- Dynamic PII redaction and virtual sandboxing
- Supports Claude, Cursor, OpenAI, LangChain, CrewAI, AutoGen, DeepSeek Harness

## Integrations
Anthropic Claude Desktop, Claude Code (MCP), Cursor IDE, VS Code, OpenAI SDK, LangChain, CrewAI, Microsoft AutoGen, DeepSeek Harness (DSH), Docker, Kubernetes, Ansible, GitHub Container Registry (GHCR)

## Platforms
WINDOWS, MACOS, LINUX, WEB, API, VSC_EXTENSION, CLI

## Pricing
Open Source, Free tier available

## Version
1.0.0

## Links
- Website: https://dr-os.io
- Documentation: https://dr-os.io/docs
- Repository: https://github.com/Top-Celestial-Company-Ltd/DROS-VajraClaw-Hacker
- EveryDev.ai: https://www.everydev.ai/tools/dros-vajraclaw-hacker