Skip to content

Harombe

Self-hosted agent framework for distributed AI with defense-in-depth security

Table of Contents

What is Harombe?

Harombe is a self-hosted AI agent framework designed for secure, distributed AI workloads. It provides a complete security layer with defense-in-depth protection, enabling you to run autonomous AI agents safely in production environments.

Status: Phase 4 Complete ✅ | Phase 5 Complete ✅ | Phase 6 Complete ✅

Key Features

Security

  • Zero-Trust Code Execution: All code runs in gVisor-isolated sandboxes with syscall filtering (70 vs 300+ syscalls)
  • Credential Security: Secrets stored in HashiCorp Vault, never in code or logs (>99% detection rate)
  • Network Isolation: Default-deny egress with domain allowlisting and private IP blocking
  • Complete Auditability: Immutable audit trail with 0.56ms write latency
  • Human-in-the-Loop: Risk-based approval gates for high-risk operations

Intelligence

  • Semantic Memory: RAG-powered context retrieval with vector embeddings
  • Multi-Modal Support: Text, voice, and browser automation
  • Tool Integration: Extensible tool system with MCP protocol support
  • Context Management: Intelligent context windowing and compression
  • Privacy Router: Hybrid local/cloud AI with PII detection and automatic routing

Performance

  • High Throughput: 601,249 operations/sec for HITL classification
  • Low Latency: <1ms for most security operations
  • Minimal Overhead: 0.32ms code execution overhead
  • Scalable: Unlimited concurrent sandboxes (CPU/memory limited)

Compliance

  • PCI DSS 4.0: Requirements 3, 6, 8, 10 compliant
  • GDPR: Articles 5, 17, 25, 30, 32, 33 compliant
  • SOC 2 Type II: CC6.1, CC6.6, CC6.7, CC7.2, CC8.1 compliant
  • NIST CSF: Identify, Protect, Detect, Respond aligned

Quick Start

# Clone repository
git clone https://github.com/smallthinkingmachines/harombe.git
cd harombe

# Install dependencies
pip install -e ".[dev]"

# Initialize (detects hardware, recommends model)
harombe init

# Pull recommended model
ollama pull qwen2.5:7b

# Start interactive agent
harombe chat

Get Started →

Architecture Overview

graph TB
    User[User/API] --> Gateway[API Gateway]
    Gateway --> Agent[Agent Runtime]

    Agent --> Memory[Memory/RAG]
    Agent --> HITL[HITL Gateway]
    Agent --> Sandbox[Sandbox Manager]

    HITL --> Vault[HashiCorp Vault]
    Sandbox --> Network[Network Filter]
    Network --> gVisor[gVisor Sandbox]

    Agent --> Audit[Audit Logger]
    Agent --> Scanner[Secret Scanner]

    Memory --> Chroma[ChromaDB]

    style Vault fill:#e8f5e9
    style gVisor fill:#e3f2fd
    style Audit fill:#fff3e0
    style HITL fill:#fce4ec

Security Layers

Harombe implements five layers of defense-in-depth security:

  1. Layer 1: Audit Logging - Immutable event trail (WAL mode, <1ms writes)
  2. Layer 2: Execution Isolation - gVisor sandbox (70 syscalls, resource limits)
  3. Layer 3: Credential Management - Vault-based secrets (no plaintext, auto-rotation)
  4. Layer 4: Network Security - Default-deny egress (allowlist, DPI)
  5. Layer 5: Human-in-the-Loop - Risk-based approvals (context-aware, auto-approval for low-risk)

Learn More →

Performance Metrics

Component Target Actual Achievement
Audit Log Write <10ms 0.56ms 17.9x faster
Code Execution <100ms 0.32ms 312x faster
HITL Classification <50ms 0.0001ms 500,000x faster
Sandbox Creation <3s 2-3s Meets target
Throughput >1K/s 601K ops/s 601x higher

HITL classification achieves 500,000x improvement because the risk classifier uses simple rule matching rather than ML inference — this is a deliberate design choice favoring determinism and speed over complexity. See the full benchmark methodology and reproduction steps for details.

Use Cases

Autonomous DevOps Agents

Harombe enables secure autonomous agents that can:

  • Execute code in isolated sandboxes
  • Access production credentials safely via Vault
  • Require human approval for high-risk operations
  • Provide complete audit trail for compliance

AI-Powered Customer Support

Build intelligent customer support agents with:

  • Semantic memory for context-aware responses
  • Multi-modal support (text, voice, browser)
  • Secure API integrations
  • Compliance-ready audit logs

Secure Code Analysis

Analyze and execute untrusted code safely:

  • gVisor isolation prevents host compromise
  • Network filtering blocks data exfiltration
  • Secret scanning prevents credential leaks
  • Complete audit trail for security review

Development Phases

  • ✅ Phase 0: Foundation (Core agent, API, tools)
  • ✅ Phase 1: Memory & Persistence (ChromaDB, semantic memory)
  • ✅ Phase 2: RAG Integration (Embeddings, retrieval)
  • ✅ Phase 3: Voice Interface (Speech-to-text, text-to-speech)
  • ✅ Phase 4: Security Layer (Sandboxing, credentials, network, audit, HITL)
  • ✅ Phase 5: Intelligence (ML anomaly detection, auto-approvals, secret rotation, privacy router)
  • ✅ Phase 6: Advanced Security (Hardware security, isolation, ZKP, distributed cryptography)

View Roadmap →

Community

License

Harombe is open source software licensed under the Apache 2.0 License.

Ready to get started? Check out the Quick Start Guide or jump into the Security Architecture.