The agent harness performance optimization system. Skills, instincts, memory, security, and research-first development for Claude Code, Codex, Opencode, Cursor and beyond.
Desktop harness for Claude Code, Codex, and ACP agents
Harnss is a desktop workspace for agent-led coding. It lets you run Claude Code, Codex, and other Agent Client Protocol agents in parallel, then inspect tool calls, diffs, MCP tools, git actions, and browser context in one app. It also keeps projects, sessions, and permissions scoped so you can switch without losing context.
Builders who want their coding agents, tools, and project state in one desktop workspace.
You can switch between agents and keep context, tool state, and project work moving without juggling terminals.
What it does
Multi-engine sessions
Runs Claude Code, Codex, and ACP-compatible agents in parallel, each with separate history and state.
Rich tool visualization
Shows tool calls as interactive cards with diffs, syntax highlighting, inline bash output, and nested subagent progress.
MCP server management
Connects MCP servers per project over stdio, SSE, or HTTP and handles OAuth in the app.
Git integration
Lets you stage, commit, push, browse branches, view history, and manage worktrees inside the app.
Built-in terminal and browser
Includes a multi-tab terminal and embedded browser that stay mounted while you work.
Agent store
Lets you browse and install ACP agents from the registry or add custom agents with command, args, env, and icon.
Plan mode and permissions
Supports plan mode plus Ask First, Accept Edits, and Allow All permission levels.
Session search and import
Searches sessions and resumes conversations started in the Claude Code CLI.
How to get it
- 1[!NOTE] Pre-built binaries are currently unsigned. On macOS, right-click the app and…
git clone https://github.com/OpenSource03/harnss.git cd harnss pnpm install pnpm dev
- 2Run
pnpm dist:mac # macOS DMG (arm64 + x64) pnpm dist:win # Windows NSIS installer (x64 + ARM64) pnpm dist:linux # Linux AppImage + deb
README
[!CAUTION] Harnss is currently pending a large rewrite to improve usability, performance and overall codebase into a more production ready app. I will share more info about it soon.
[!WARNING] Harnss is in early development and issues are to be expected. Please feel free to report bugs and issues in the Issues section.
Harness your AI coding agents.
Harnss is a cross-platform desktop app that gives you one interface to run, manage, and switch between AI coding agents — Claude Code, Codex, and any ACP-compatible agent — without losing context, sessions, or tool state.
Why Harnss?
- One app, every agent. Run Claude Code, Codex, and custom ACP agents side by side. No more juggling terminals or losing context when switching tools.
- See what your AI is actually doing. Tool calls render as interactive cards with word-level diffs, syntax highlighting, and inline bash output — not raw JSON.
- Your workspace, your way. Built-in terminal, browser, git, MCP servers, and file panels — all scoped per project, all staying open while you work.
Screenshots
Terminal, browser, git, and MCP integrations — all in one window.
One UI - all agents. Browse entire ACP registry.
Click to reveal other screenshots
https://github.com/user-attachments/assets/41f5bcbc-c141-4a16-b430-79f8e01d107e
Organize your projects in spaces.
Every tool call beautifully visualised - even those from popular MCPs.
Run multiple agent sessions side by side — switch instantly without losing progress.
Features
Multi-engine sessions
Run Claude Code (via the Anthropic SDK), Codex, and ACP-compatible agents in parallel. Each session has its own state, history, and context. Switch between them instantly.
Rich tool visualization
Every tool call renders as an interactive card. File edits show word-level diffs with syntax highlighting. Bash output appears inline. Subagent tasks nest with step-by-step progress tracking. File changes are summarized per turn with a dedicated Changes panel.
MCP server management
Connect any MCP server per project via stdio, SSE, or HTTP transport. OAuth flows are handled automatically. Server status and available tool counts are visible at a glance. Jira, Confluence, and other integrations render with dedicated UIs rather than raw JSON.
Git integration
Stage, unstage, commit, and push without leaving the app. Browse branches, view commit history, and manage git worktrees. AI-generated commit messages are available from the staged diff.
Built-in terminal & browser
Multi-tab PTY terminal backed by native shell processes. An embedded browser for opening URLs inline and providing additional context to the agent. Both panels stay mounted while you work.
Project workspaces & spaces
Projects map to folders on disk. Spaces let you organize projects into named groups with custom icons and colors. Sessions, history, and panel settings are all scoped per project.
Agent Store
Browse and install agents from the ACP community registry directly in the app. Add custom agents by specifying a command, arguments, environment variables, and an icon. All configuration is managed through Settings — no config files.
Plan mode & permission control
Work in plan mode to have the agent draft a plan before making any changes. Three permission levels — Ask First, Accept Edits, Allow All — control how much autonomy the agent has. Switch modes at any point without interrupting context.
Background task agents
Task agents spawned during a session continue running in the background and are tracked in a dedicated panel. Keep working in other sessions while long-running tasks complete.
Image attachments & annotation
Attach screenshots or images directly in the chat. An built-in annotation tool lets you draw, highlight, and mark up images with freehand strokes before sending them to the agent.
Voice input & notifications
Voice input via native macOS dictation or an on-device Whisper model (no API key required). Configurable OS notifications for plan approval requests, permission prompts, agent questions, and session completion.
Session search & history
Full-text search across session titles and message content. Import and resume conversations previously started in the Claude Code CLI.
Quick Start
- Download the latest release for your platform from the Releases page
- Open a project — point Harnss at any folder on disk
- Choose an engine — Claude Code, Codex, or any installed ACP agent — and start working
Engines & Agents
Harnss supports three execution engines out of the box:
| Engine | Protocol | Requirements |
|---|---|---|
| Claude Code | Anthropic Agent SDK | Claude account (subscription or API key) |
| Codex | JSON-RPC app-server | Codex CLI in PATH + OpenAI API key or ChatGPT account |
| ACP agents | Agent Client Protocol | Agent-specific (see registry) |
Claude Code and Codex are built-in. ACP agents can be installed from the ACP Agent Registry inside the app, or configured manually.
Examples of installable ACP-compatible agents:
| Agent | Command | Notes |
|---|---|---|
| Gemini CLI | gemini --experimental-acp | Experimental ACP flag |
| Goose | goose acp | |
| Docker cagent | cagent acp agent.yml | Container-based agents |
Adding an agent
Open Settings → ACP Agents. The Agent Store tab lets you browse and install agents from the community registry. The My Agents tab lets you create custom agents — set the binary command, arguments, environment variables, and icon, or paste a JSON definition to auto-fill the form.
MCP Servers
MCP servers are configured per project through the MCP Servers panel in the right-side toolbar. Supported transports: stdio, SSE, and HTTP. OAuth authentication is handled in-app with token persistence across sessions.
Install
[!NOTE] Pre-built binaries are currently unsigned. On macOS, right-click the app and choose Open to bypass the Gatekeeper warning on first launch. On Windows, click More info → Run anyway if Windows Defender flags the installer.
| Platform | Download |
|---|---|
| macOS (Apple Silicon) | .dmg (arm64) |
| macOS (Intel) | .dmg (x64) |
| Windows (x64) | .exe installer |
| Windows (ARM64) | .exe installer |
| Linux | .AppImage / .deb |
Development
git clone https://github.com/OpenSource03/harnss.git
cd harnss
pnpm install
pnpm dev
Build installers
pnpm dist:mac # macOS DMG (arm64 + x64)
pnpm dist:win # Windows NSIS installer (x64 + ARM64)
pnpm dist:linux # Linux AppImage + deb
Contributing
- Fork the repo and create a feature branch
- Follow the conventions in
CLAUDE.md - Test with
pnpm dev - Open a pull request
License
MIT
Built on the Agent Client Protocol
Files in the repo
- .agents
- .claude
- .github
- .harnss
- build
- electron
- public
- scripts
- shared
- src
- .gitignore
- .npmrc
- AGENTS.md
- CLAUDE.md
- components.json
- electron-builder.config.js
- index.html
- LICENSE
- package.json
- pnpm-lock.yaml
- pnpm-workspace.yaml
- README.md
- skills-lock.json
- tsconfig.json
- tsup.electron.config.ts
- vite.config.ts
- vitest.config.electron.ts
Discussion (0)
Ask about usage, or say what you built with itSign in to join the discussion.
No comments yet. Be the first to say what this is good for.
More harnesses
from vibe coding to agentic engineering - practice makes claude perfect
🌊 The original agent meta-harness. Deploy intelligent multi-player swarms, coordinate autonomous workflows, and build conversational AI systems. Features adaptive memory, self-learning intelligence, RAG integration, and native Claude Code / Codex / Hermes and many more Integrated
Practical patterns, starters & CLI tools for loop engineering with AI coding agents. Design systems that prompt and orchestrate agents (inspired by Addy Osmani and Boris Cherny). Includes loop-audit, loop-init, loop-cost.
Git. Ship. Done - Core

The most RAM efficient harness