Run and supervise AI coding agents on the machines where your code already lives, from anywhere.
Harness Remote is a local-first control plane for AI coding agents. Connect to the machine where your code and credentials already live, then supervise OpenCode, Claude Code, Codex CLI, Oh My Pi and PI from one interface on phone, web or desktop.
One interface. Multiple agents. Your machines, your credentials, your code.
Harness Remote is not another coding agent. It is the control plane above them.
Harness Remote
phone · web · desktop
│
┌────────┴────────┐
│ Machine Daemon │
└────────┬────────┘
┌────────────┼────────────┐
Codex Claude OpenCode
OMP PI ...
Execution stays on your machines. Repositories stay on your machines. Agent credentials and model access stay on your machines. Harness Remote coordinates and supervises the work remotely.
On the machine where your coding agents are installed:
npx github:giuliastro/harness-remoteIt detects the supported agents on your PATH, picks a free port, generates credentials and prints the address to enter in the client. OpenCode is started and supervised directly; ACP-backed agents run through the local bridge or daemon.
Then install a client from GitHub Releases, or open the web app and enter the address it printed.
Harness Remote already gives you a common remote UI for five coding-agent harnesses:
| Harness | Support |
|---|---|
| OpenCode | direct HTTP + managed daemon host |
| Claude Code | ACP bridge |
| Codex CLI | ACP bridge |
| Oh My Pi (OMP) | ACP bridge |
| PI | ACP bridge |
From Android, the web/PWA or the desktop app you can monitor sessions, read streamed progress, send prompts, stop work, select models where supported, inspect agent questions/todos and use the capabilities each harness exposes.
The Harness 3 daemon is now able to represent multiple agent hosts under one stable machine identity and route requests through a single machine connection. Legacy per-harness connections remain supported while the client UI moves to machine-first agent discovery and selection.
The goal is not just remote chat with coding agents. The goal is a local-first operating layer for coding work across agents and machines:
- Attention plane — one place for the questions, permissions, failures and completed work that actually need you.
- Multi-agent machine daemon — one machine connection, multiple heterogeneous agents, independent lifecycle and failure isolation.
- Task-oriented execution — launch work against a repository instead of manually preparing every agent session.
- Worktree isolation — concurrent tasks get isolated Git workspaces rather than colliding in one checkout.
- Finish-work loop — diff → tests → review → PR as a first-class workflow.
- Multi-machine fleet — supervise and eventually route work across desktops, laptops and servers without moving credentials or repositories into a hosted control plane.
See docs/HARNESS_3_ROADMAP.md for the architecture and implementation roadmap.
Coding agents are most useful where the repositories, build tools, local models, subscriptions and credentials already are. Harness Remote keeps that execution boundary intact and adds the missing remote control layer on top.
That means you can leave a workstation or server doing the work while you use another device to check progress, answer the agent, stop a bad run or start the next step.
- Android — native Capacitor app.
- Web / PWA — installable web client published from this repository.
- Desktop — Electron builds for Windows, macOS and Linux.
Current screenshots:
| Sessions | Detail |
|---|---|
 |
 |
The previous full README — including detailed OpenCode, OMP, PI, Claude Code and Codex setup, Android/Desktop/PWA notes, security caveats, endpoints and build instructions — is preserved as REFERENCE.md.
Other useful docs:
Harness Remote is evolving from a multi-harness remote client into a local-first control plane. The repository deliberately keeps backward compatibility while that transition lands in small, reviewable slices.
If that is a problem you have too — several coding agents, several machines, and too much terminal babysitting — issues and feedback are especially useful now.