HKUDS/OpenHarness

  
oh — OpenHarness & ohmo

English · 简体中文

OpenHarness delivers core lightweight agent infrastructure: tool-use, skills, memory, and multi-agent coordination.

ohmo is a personal AI agent built on OpenHarness — not another chatbot, but an assistant that actually works for you over long sessions. Chat with ohmo in Feishu / Slack / Telegram / Discord, and it forks branches, writes code, runs tests, and opens PRs on its own. ohmo runs on your existing Claude Code or Codex subscription — no extra API key needed.

Join the community: contribute Harness for open agent development.

One Command (oh) to Launch OpenHarness and Unlock All Agent Harnesses.

Supports CLI agent integration including OpenClaw, nanobot, Cursor, and more.

✨ OpenHarness's Key Harness Features

🔄 Agent Loop • Streaming Tool-Call Cycle • API Retry with Exponential Backoff • Parallel Tool Execution • Token Counting & Cost Tracking

🔧 Harness Toolkit • 43 Tools (File, Shell, Search, Web, MCP) • On-Demand Skill Loading (.md) • Plugin Ecosystem (Skills + Hooks + Agents) • Compatible with anthropics/skills & plugins

🧠 Context & Memory • CLAUDE.md Discovery & Injection • Context Compression (Auto-Compact) • MEMORY.md Persistent Memory • Session Resume & History

🛡️ Governance • Multi-Level Permission Modes • Path-Level & Command Rules • PreToolUse / PostToolUse Hooks • Interactive Approval Dialogs

🤝 Swarm Coordination • Subagent Spawning & Delegation • Team Registry & Task Management • Background Task Lifecycle • ClawTeam Integration (Roadmap)

🤔 What is an Agent Harness?

An Agent Harness is the complete infrastructure that wraps around an LLM to make it a functional agent. The model provides intelligence; the harness provides hands, eyes, memory, and safety boundaries.

OpenHarness is an open-source Python implementation designed for researchers, builders, and the community:

• Understand how production AI agents work under the hood
• Experiment with cutting-edge tools, skills, and agent coordination patterns
• Extend the harness with custom plugins, providers, and domain knowledge
• Build specialized agents on top of proven architecture

📰 What's New

• Unreleased 🔍 Dry-run safe preview:
  • oh --dry-run previews resolved runtime settings, auth state, skills, commands, tools, and configured MCP servers without executing the model, tools, or subagents.
  • Dry-run now reports a ready / warning / blocked readiness verdict with concrete next-step suggestions such as fixing auth, fixing MCP config, or running the prompt directly.
  • Prompt previews include likely matching skills and tools, while slash-command previews show whether the command is mostly read-only or stateful.
• 2026-04-18 ⚙️ v0.1.7 — Packaging & TUI polish:
  • Install script now links oh, ohmo, and openharness into ~/.local/bin instead of prepending the virtualenv bin directory to PATH, which avoids clobbering Conda-managed shells.
  • React TUI now supports Shift+Enter to insert a newline while keeping plain Enter as submit.
  • Busy-state animation in the React TUI is quieter and less error-prone on Windows terminals, with conservative spinner frames and reduced flashing.
• 2026-04-10 🧠 v0.1.6 — Auto-Compaction & Markdown TUI:
  • Auto-Compaction preserves task state and channel logs across context compression — agents can run multi-day sessions without manual compact/clear
  • Subprocess teammates run in headless worker mode; agent team creation stabilized
  • Assistant messages now render full Markdown in the React TUI
  • ohmo gains channel slash commands and multimodal attachment support
• 2026-04-08 🔌 v0.1.5 — MCP HTTP transport & Swarm polling:
  • MCP protocol adds HTTP transport, auto-reconnect on disconnect, and tool-only server compatibility
  • JSON Schema types inferred for MCP tool inputs — no manual type mapping needed
  • ohmo channels support file attachments and multimodal gateway messages
  • Subprocess agents are now pollable in real runs; permission modals serialized to prevent input swallowing
• 2026-04-08 🌙 v0.1.4 — Multi-provider auth & Moonshot/Kimi:
  • Native Moonshot/Kimi provider with reasoning_content support for thinking models
  • Auth overhaul: fixed provider-switching key mismatch, OPENAI_BASE_URL env override, profile-scoped credential priority
  • MCP gracefully handles disconnected servers in call_tool / read_resource
  • Security: built-in sensitive-path protection in PermissionChecker, hardened web_fetch URL validation
  • Stability: EIO crash recovery in Ink TUI, --debug logging, Windows cmd flash fix
• 2026-04-06 🚀 v0.1.2 — Unified setup flows and ohmo personal-agent app:
  • oh setup now guides provider selection as workflows instead of exposing raw auth/provider internals
  • Compatible API setup is now profile-scoped, so Anthropic/OpenAI-compatible endpoints can keep separate keys
  • ohmo ships as a packaged app with ~/.ohmo workspace, gateway, bootstrap prompts, and channel config flow
• 2026-04-01 🎨 v0.1.0 — Initial OpenHarness open-source release featuring complete Harness architecture:

Start here: Quick Start · Provider Compatibility · Showcase · Contributing · Changelog

🚀 Quick Start

1. Install

Linux / macOS / WSL

# One-click install
curl -fsSL https://raw.githubusercontent.com/HKUDS/OpenHarness/main/scripts/install.sh | bash

# Or via pip
pip install openharness-ai

Windows (Native)

# One-click install (PowerShell)
iex (Invoke-WebRequest -Uri 'https://raw.githubusercontent.com/HKUDS/OpenHarness/main/scripts/install.ps1')

# Or via pip
pip install openharness-ai

Note: Windows support is now native. In PowerShell, use openh instead of oh because oh can resolve to the built-in Out-Host alias.

2. Configure

oh setup    # interactive wizard — pick a provider, authenticate, done
# On Windows PowerShell, use: openh setup

Supports Claude / OpenAI / Copilot / Codex / Moonshot(Kimi) / GLM / MiniMax / NVIDIA NIM and any compatible endpoint.

3. Run

oh
# On Windows PowerShell, use: openh

4. Set up ohmo (Personal Agent)

Want an AI agent that works for you from Feishu / Slack / Telegram / Discord?

ohmo init             # initialize ~/.ohmo workspace
ohmo config           # configure channels and provider
ohmo gateway start    # start the gateway — ohmo is now live in your chat app

ohmo runs on your existing Claude Code subscription or Codex subscription — no extra API key needed.

Non-Interactive Mode (Pipes & Scripts)

# Single prompt → stdout
oh -p "Explain this codebase"

# JSON output for programmatic use
oh -p "List all functions in main.py" --output-format json

# Stream JSON events in real-time
oh -p "Fix the bug" --output-format stream-json

Dry Run (Safe Preview)

Use --dry-run when you want to inspect what OpenHarness would use before any live execution starts.

# Preview an interactive session setup
oh --dry-run

# Preview one prompt without executing the model or tools
oh --dry-run -p "Review this bug fix and grep for failing tests"

# Preview a slash command path
oh --dry-run -p "/plugin list"

# Get structured output for scripts or channels
oh --dry-run -p "Explain this repository" --output-format json

Dry-run is intentionally static:

• It does not call the model
• It does not execute tools or spawn subagents
• It does not connect to MCP servers
• It does resolve settings, auth status, prompt assembly, skills, commands, tools, and obvious MCP config problems

Readiness levels:

• ready: configuration looks usable; the next suggested action is usually to run the prompt directly
• warning: OpenHarness can resolve the session, but something important still looks wrong, such as broken MCP config or missing auth for later model work
• blocked: the requested path will not run successfully as-is, for example an unknown slash command or a prompt that cannot resolve a runtime client

next actions in the dry-run output tell you the shortest fix or follow-up step, such as:

• run oh auth login
• fix or disable broken MCP configuration
• run the prompt directly with oh -p "..." or open the interactive UI with oh

🔌 Provider Compatibility

OpenHarness treats providers as workflows backed by named profiles. In day-to-day use, prefer:

oh setup
oh provider list
oh provider use <profile>

Built-in Workflows

Compatible API Families

Anthropic-Compatible API

Typical examples:

OpenAI-Compatible API

Any provider implementing the OpenAI /v1/chat/completions style API works:

Advanced Profile Management

# List saved workflows
oh provider list

# Switch the active workflow
oh provider use codex

# Add your own compatible endpoint
oh provider add my-endpoint \
  --label "My Endpoint" \
  --provider openai \
  --api-format openai \
  --auth-source openai_api_key \
  --model my-model \
  --base-url https://example.com/v1

For custom compatible endpoints, OpenHarness can bind credentials per profile instead of forcing every Anthropic-compatible or OpenAI-compatible backend to share the same API key.

Ollama (Local Models)

Run local models through Ollama's OpenAI-compatible endpoint:

# Add an Ollama provider profile
oh provider add ollama \
  --label "Ollama" \
  --provider Ollama \
  --api-format openai \
  --auth-source openai_api_key \
  --model glm-4.7-flash:q8_0 \
  --base-url http://localhost:11434/v1

Saved provider profile: ollama

# Activate and verify
oh provider use ollama

Activated provider profile: ollama

oh provider list

  claude-api: Anthropic-Compatible API [ready]
  ...
  moonshot: Moonshot (Kimi) [missing auth]
    auth=moonshot_api_key model=kimi-k2.5 base_url=https://api.moonshot.cn/v1
* ollama: Ollama [ready]
    auth=openai_api_key model=glm-4.7-flash:q8_0 base_url=http://localhost:11434/v1

GitHub Copilot Format (--api-format copilot)

Use your existing GitHub Copilot subscription as the LLM backend. Authentication uses GitHub's OAuth device flow — no API keys needed.

# One-time login (opens browser for GitHub authorization)
oh auth copilot-login

# Then launch with Copilot as the provider
uv run oh --api-format copilot

# Or via environment variable
export OPENHARNESS_API_FORMAT=copilot
uv run oh

# Check auth status
oh auth status

# Remove stored credentials
oh auth copilot-logout

🏗️ Harness Architecture

OpenHarness implements the core Agent Harness pattern with 10 subsystems:

openharness/
  engine/          # 🧠 Agent Loop — query → stream → tool-call → loop
  tools/           # 🔧 43 Tools — file I/O, shell, search, web, MCP
  skills/          # 📚 Knowledge — on-demand skill loading (.md files)
  plugins/         # 🔌 Extensions — commands, hooks, agents, MCP servers
  permissions/     # 🛡️ Safety — multi-level modes, path rules, command deny
  hooks/           # ⚡ Lifecycle — PreToolUse/PostToolUse event hooks
  commands/        # 💬 54 Commands — /help, /commit, /plan, /resume, ...
  mcp/             # 🌐 MCP — Model Context Protocol client
  memory/          # 🧠 Memory — persistent cross-session knowledge
  tasks/           # 📋 Tasks — background task management
  coordinator/     # 🤝 Multi-Agent — subagent spawning, team coordination
  prompts/         # 📝 Context — system prompt assembly, CLAUDE.md, skills
  config/          # ⚙️ Settings — multi-layer config, migrations
  ui/              # 🖥️ React TUI — backend protocol + frontend

The Agent Loop

The heart of the harness. One loop, endlessly composable:

while True:
    response = await api.stream(messages, tools)
    
    if response.stop_reason != "tool_use":
        break  # Model is done
    
    for tool_call in response.tool_uses:
        # Permission check → Hook → Execute → Hook → Result
        result = await harness.execute_tool(tool_call)
    
    messages.append(tool_results)
    # Loop continues — model sees results, decides next action

The model decides what to do. The harness handles how — safely, efficiently, with full observability.

Harness Flow

flowchart LR
    U[User Prompt] --> C[CLI or React TUI]
    C --> R[RuntimeBundle]
    R --> Q[QueryEngine]
    Q --> A[Anthropic-compatible API Client]
    A -->|tool_use| T[Tool Registry]
    T --> P[Permissions + Hooks]
    P --> X[Files Shell Web MCP Tasks]
    X --> Q

✨ Features

🔧 Tools (43+)

Every tool has:

• Pydantic input validation — structured, type-safe inputs
• Self-describing JSON Schema — models understand tools automatically
• Permission integration — checked before every execution
• Hook support — PreToolUse/PostToolUse lifecycle events

📚 Skills System

Skills are on-demand knowledge — loaded only when the model needs them:

Available Skills:
- commit: Create clean, well-structured git commits
- review: Review code for bugs, security issues, and quality
- debug: Diagnose and fix bugs systematically
- plan: Design an implementation plan before coding
- test: Write and run tests for code
- simplify: Refactor code to be simpler and more maintainable
- pdf: PDF processing with pypdf (from anthropics/skills)
- xlsx: Excel operations (from anthropics/skills)
- ... 40+ more

Skills can live in bundled, user, ohmo, project, or plugin locations. User-level skills are loaded from:

~/.openharness/skills/<skill>/SKILL.md
~/.claude/skills/<skill>/SKILL.md
~/.agents/skills/<skill>/SKILL.md

Project-level skills are enabled by default and are discovered from the current working directory up to the git root:

<project>/.openharness/skills/<skill>/SKILL.md
<project>/.agents/skills/<skill>/SKILL.md
<project>/.claude/skills/<skill>/SKILL.md

Disable project skills for untrusted repositories with:

oh config set allow_project_skills false

Use /skills to list loaded skills with their source and path. User-invocable skills can be run directly as slash commands, for example /deploy staging.

Compatible with anthropics/skills — use the SKILL.md directory layout above.

🌐 Web search and proxy settings

Built-in web_search uses DuckDuckGo HTML search by default. In regions where that endpoint is unreachable, point OpenHarness at a trusted public HTML search endpoint or your own SearXNG instance:

export OPENHARNESS_WEB_SEARCH_URL="https://your-searxng.example/search"

web_search and web_fetch keep trust_env=False for SSRF safety, so they do not automatically inherit HTTP_PROXY / HTTPS_PROXY. If you need a proxy, opt in with an OpenHarness-specific variable:

export OPENHARNESS_WEB_PROXY="http://127.0.0.1:7890"

The proxy URL must be HTTP/HTTPS and cannot contain embedded credentials.

🔌 Plugin System

Compatible with claude-code plugins. Tested with 12 official plugins:

# Manage plugins
oh plugin list
oh plugin install <source>
oh plugin enable <name>

🤝 Ecosystem Workflows

OpenHarness is useful as a lightweight harness layer around Claude-style tooling conventions:

• OpenClaw-oriented workflows can reuse Markdown-first knowledge and command-driven collaboration patterns.
• Claude-style plugins and skills stay portable because OpenHarness keeps those formats familiar.
• ClawTeam-style multi-agent work maps well onto the built-in team, task, and background execution primitives.

For concrete usage ideas instead of generic claims, see docs/SHOWCASE.md.

🛡️ Permissions

Multi-level safety with fine-grained control:

Path-level rules in settings.json:

{
  "permission": {
    "mode": "default",
    "path_rules": [{"pattern": "/etc/*", "allow": false}],
    "denied_commands": ["rm -rf /", "DROP TABLE *"]
  }
}

🖥️ Terminal UI

React/Ink TUI with full interactive experience:

• Command picker: Type / → arrow keys to select → Enter
• Permission dialog: Interactive y/n with tool details
• Mode switcher: /permissions → select from list
• Session resume: /resume → pick from history
• Animated spinner: Real-time feedback during tool execution
• Keyboard shortcuts: Shown at the bottom, context-aware

📡 CLI

oh [OPTIONS] COMMAND [ARGS]

Session:     -c/--continue, -r/--resume, -n/--name
Model:       -m/--model, --effort, --max-turns
Output:      -p/--print, --output-format text|json|stream-json
Permissions: --permission-mode, --dangerously-skip-permissions
Context:     -s/--system-prompt, --append-system-prompt, --settings
Advanced:    -d/--debug, --mcp-config, --bare

Subcommands: oh setup | oh provider | oh auth | oh mcp | oh plugin

🧑‍💼 ohmo Personal Agent

ohmo is a personal-agent app built on top of OpenHarness. It is packaged alongside oh, with its own workspace and gateway:

# Initialize personal workspace
ohmo init

# Configure gateway channels and pick a provider profile
ohmo config

# Run the personal agent
ohmo

# Run the gateway in foreground
ohmo gateway run

# Check or restart the gateway
ohmo gateway status
ohmo gateway restart

Key concepts:

• ~/.ohmo/
  • personal workspace root
• soul.md
  • long-term agent personality and behavior
• identity.md
  • who ohmo is
• user.md
  • user profile and preferences
• BOOTSTRAP.md
  • first-run landing ritual
• memory/
  • personal memory
• gateway.json
  • selected provider profile and channel configuration

ohmo config uses the same workflow language as oh setup, so you can point the personal-agent gateway at:

• Anthropic-Compatible API
• Claude Subscription
• OpenAI-Compatible API
• Codex Subscription
• GitHub Copilot

ohmo init creates the home workspace once. After that, use ohmo config to update provider and channel settings; if the gateway is already running, the config flow can restart it for you.

Currently ohmo init / ohmo config can guide channel setup for:

• Telegram
• Slack
• Discord
• Feishu

📊 Test Results

# Run all tests
uv run pytest -q                           # 114 unit/integration
python scripts/test_harness_features.py     # Harness E2E
python scripts/test_real_skills_plugins.py  # Real plugins E2E

🔧 Extending OpenHarness

Add a Custom Tool

from pydantic import BaseModel, Field
from openharness.tools.base import BaseTool, ToolExecutionContext, ToolResult

class MyToolInput(BaseModel):
    query: str = Field(description="Search query")

class MyTool(BaseTool):
    name = "my_tool"
    description = "Does something useful"
    input_model = MyToolInput

    async def execute(self, arguments: MyToolInput, context: ToolExecutionContext) -> ToolResult:
        return ToolResult(output=f"Result for: {arguments.query}")

Add a Custom Skill

Create ~/.openharness/skills/my-skill.md:

---
name: my-skill
description: Expert guidance for my specific domain
---

# My Skill

## When to use
Use when the user asks about [your domain].

## Workflow
1. Step one
2. Step two
...

Add a Plugin

Create .openharness/plugins/my-plugin/.claude-plugin/plugin.json:

{
  "name": "my-plugin",
  "version": "1.0.0",
  "description": "My custom plugin"
}

Add commands in commands/*.md, hooks in hooks/hooks.json, agents in agents/*.md.

🌍 Showcase

OpenHarness is most useful when treated as a small, inspectable harness you can adapt to a real workflow:

• Repo coding assistant for reading code, patching files, and running checks locally.
• Headless scripting tool for json and stream-json output in automation flows.
• Plugin and skill testbed for experimenting with Claude-style extensions.
• Multi-agent prototype harness for task delegation and background execution.
• Provider comparison sandbox across Anthropic-compatible backends.

See docs/SHOWCASE.md for short, reproducible examples.

🤝 Contributing

OpenHarness is a community-driven research project. We welcome contributions in:

# Development setup
git clone https://github.com/HKUDS/OpenHarness.git
cd OpenHarness
uv sync --extra dev
uv run pytest -q  # Verify everything works

Useful contributor entry points:

• CONTRIBUTING.md for setup, checks, and PR expectations
• CHANGELOG.md for user-visible changes
• docs/SHOWCASE.md for real-world usage patterns worth documenting

🔧 Troubleshooting

Backspace key in macOS Terminal.app

OpenHarness handles both common terminal delete sequences, including the raw DEL byte (0x7f) that macOS Terminal.app sends for Backspace. If Backspace inserts spaces or visible control characters instead of deleting text, upgrade OpenHarness first.

For older versions that do not include this fix, use a terminal that sends a standard Backspace sequence or adjust your terminal keyboard profile as a temporary workaround.

📄 License

MIT — see LICENSE.

Oh my Harness!
The model is the agent. The code is the harness.

Thanks for visiting ✨ OpenHarness!