OpenHarness Daemon v1.7.0 • Production Ready

The Autonomous,
Multi-Runtime AI Daemon

A resilient, long-lived autonomous agent platform engineered in Go. Combines persistent markdown memory, hybrid BM25 + FVEC vector retrieval, sandboxed Docker multi-runtimes, and a decentralized peer agent mesh.

Get Started in 30 Seconds Live Daemon Simulator Browse GitHub

bash - openharness quickstart

# Quick install pre-compiled binary (Linux & macOS)
curl -fsSL https://openharness.dev/install.sh | bash

# Verify installation
openharness --version
# Output: openharness v1.7.0 (built with go1.26 darwin/arm64)

⚡ Pure Go Core • < 30MB RAM Footprint • Zero Runtime Dependencies

🔒 Non-Root Sandbox Security

Interactive Execution Engine

5-Tier Priority Inbox & Autonomous Loop

Experience how OpenHarness prioritizes high-criticality container alerts over collaborator chats, subagent completions, peer mesh messages, and webhooks.

Event Injector Interactive

Trigger incoming events into the priority queues to watch the agent loop respond.

P0 Queue 0

P1 Queue 0

P2 Queue 0

Domain Agent Loop Stream

Daemon Idle (Listening)

Tokens: 4,280 / 32,000

21:50:00 LOOP OpenHarness daemon started. BM25 and FVEC vector indexes reconciled. ready

21:50:02 P1 MSG Telegram adapter connected @bot_openharness

Hexagonal Domain Decoupled • Autonomous Prompt Evolution • Zero-Loss Graceful State Save

Core Architecture

Built for Relentless Autonomy

Engineered from scratch in Go to survive network interruptions, LLM context limits, and long-running daemon execution.

Persistent Markdown Memory & Hybrid Dual Search

In-Memory BM25 + Binary FVEC v1 Semantic Embeddings

All agent memories are stored in human-readable Markdown with full subfolder hierarchy and soft-delete .trash/ safety. OpenHarness performs dual-engine search: BM25 lexical ranking for exact symbols and paragraph-aligned cosine vector similarity for semantic discovery.

// Dual section search result format

## Lexical Matches (BM25)

- memory/architecture/sandbox.md (score: 14.2)

## Semantic Matches (FVEC v1)

- memory/notes/docker-nonroot.md#3 (similarity: 0.892)

Arch Multi-Runtime Sandbox

Python (uv), Node, Bun, Deno, Go

Hardened non-root ephemeral containers. The agent runs shell pipelines, installs packages on the fly, and runs code across multiple language runtimes without polluting the host.

uv / python node / bun / deno go 1.26 git + gh CLI

Container Daemons

daemon_spawn & P0 Alerts

The agent can spawn long-running background services (log watchers, metric monitors, web scrapers) that stream high-priority P0 events directly into the daemon inbox.

Agent Skills & Security Audit

agentskills.io Standard + AI Audit

Autonomously install Agent Skills from URLs or Git repos. Before any skill executes, an isolated security subagent audits the code for exfiltration and credential leaks.

Subagents & Peer Mesh

Hierarchical & P2P Agent Mesh

Delegate specialized tasks to synchronous or asynchronous subagents. Connect multiple OpenHarness daemons across networks using peer-to-peer authenticated messaging.

Dynamic 2-Tier Tool Catalog

Explore 25+ Built-in Tools

Tier 1 core tools are always loaded in context. Specialized Tier 2 tools are discovered dynamically via semantic tool search.

18 tools

Software Engineering

Hexagonal Architecture (Ports & Adapters)

The domain agent loop is completely isolated from external dependencies. Any LLM provider, storage backend, or chat network can be swapped without modifying agent core logic.

Inbound Ports

Operator & Mesh Adapters

Telegram Bot, Discord Bot, Webhook Listeners, and P2P Agent Mesh Inboxes feed into the 5-Tier Inbox.

LLM Port

ChatModel & Tokenizer

OpenAI API, KoboldCpp native tokencount/abort, Local Ollama, and Anthropic-compatible endpoints.

Internal Domain Hexagon

Agent Loop & State Machine

internal/agent is 100% pure Go with no I/O dependencies. Drives context compaction, idle detection, prompt migrations, and tool dispatching.

Domain Isolation: Enforced by Go internal/

Compaction: Self-Saving Summaries

Graceful Save: 2-Phase SIGINT Catch

Execution Port

Docker Multi-Runtime Sandbox

Non-root ephemeral containers executing Python uv, Node, Bun, Deno, and Go scripts in isolation.

Storage Port

Memory, BM25 & FVEC Store

Filesystem Markdown store with .trash/ soft-delete, in-memory BM25 index, and persisted binary vector embeddings.

Interactive Configurator

Generate Your Custom `config.toml`

Toggle backends, sandboxes, operators, and MCP servers to get a validated, production-ready configuration file in real time.

LLM Backend Provider

Model Identifier

Features & Adapters Docker Multi-Runtime Sandbox (Python, Node, Go) FVEC v1 Semantic Vector Search Telegram Bot Collaborator Discord Bot Collaborator Model Context Protocol (MCP) Integration Hierarchical Subagents Orchestration Web Admin UI (HTMX + DaisyUI Dashboard)

config.toml Preview

Production Ready

Deploy OpenHarness Anywhere

Run OpenHarness natively with systemd, deploy containerized via Docker Compose, and expose dashboards securely via Cloudflare Tunnel.

Native Host (Recommended)

systemd User Service

Provides direct control over Docker sandboxes on the host without socket-mounting security complications.

[Unit]
Description=OpenHarness Daemon
After=network.target

[Service]
ExecStart=/usr/local/bin/openharness -config %h/.config/openharness/config.toml
Restart=always
RestartSec=5s

[Install]
WantedBy=default.target

Auto-restarts & log rotation built-in

Containerized

Docker Compose Stack

Spin up the daemon along with pre-warmed sandbox runtimes and local vector storage.

services:
  openharness:
    image: ghcr.io/camilovalderruten/openharness:latest
    volumes:
      - ./config.toml:/etc/openharness/config.toml:ro
      - ./data:/var/lib/openharness/data
      - /var/run/docker.sock:/var/run/docker.sock
    restart: unless-stopped

Includes multi-runtime sandbox image

Zero-Trust Network

Cloudflare Tunnel

Map the local HTML Publisher and Admin Dashboard to public HTTPS domains securely without opening ports.

# cloudflared tunnel ingress
ingress:
  - hostname: agent.yourdomain.com
    service: http://localhost:8080
  - hostname: publish.yourdomain.com
    service: http://localhost:8081
  - service: http_status:404

Seamless edge TLS termination

Why OpenHarness?

Architecture Comparison

See how OpenHarness compares to standard Python agent frameworks and raw LLM scripts.

Feature	OpenHarness Daemon	LangChain / AutoGen	Raw Python Agent Scripts
Runtime Architecture	Pure Go Hexagonal Daemon	Heavy Python Framework	Ad-hoc Script Loop
RAM Consumption	< 30 MB	300 MB – 1.5 GB	150 MB – 500 MB
Long-term Memory	Markdown + BM25 + FVEC v1 Vector	External Vector DB Dependency	In-memory / Lost on restart
Code Sandbox Security	Hardened Non-Root Arch Docker (uv, bun, go)	Basic Python REPL / Docker	Unsafe Host Subprocess Execution
Multi-Agent Mesh	Built-in P2P Network Protocol	Single-process multi-agent only	None
Self-Modifying Prompts	Idempotent Prompt Migrations	Manual Code Refactor	Static Prompts

Start Building Autonomous Daemons

Deploy your first OpenHarness instance in under a minute with pre-compiled binaries or Docker.

Star on GitHub Build Your Config

The Autonomous, Multi-Runtime AI Daemon