cross-harness

Why cross-harness

Every coding agent comes with its own terminal, its own permission prompts and its own idea of the repository. Use two at once and you are copying answers between windows, approving the same command twice and untangling edits made to the same files.

cross-harness puts them on one stage. It starts each agent through the protocol its vendor ships, so it sees every command, edit, plan step and permission request as data. You address agents with @mentions, they can ask each other questions through cross-harness, everything they want to do waits in one queue for you, each works on its own git branch, and nothing is merged until a check you trust — a command cross-harness runs itself, or a review by a different agent — has passed.

cross-harness is the open cross-harness protocol from the team behind the BizOS CLI, with this reference implementation: a terminal UI, a local web UI and a command for scripts, in plain Node.js with no dependencies. The BizOS CLI is built in: type @bizos and it works next to Claude Code, Codex and the rest.

[!TIP] BizOS CLI, built in. The BizOS CLI, BizOS Code, is the BizOS team's coding agent for the terminal: one terminal, every coding agent, orchestrated. It turns one request into a dependency-aware workflow and hands each step to a native BizOS agent or a coding CLI already on your computer — Claude Code, Codex, Kimi Code, DeepSeek and more. Independent steps run in parallel, writers take turns with the workspace, and every result comes back to one conversation. Install it from code.bizos-ai.com; in cross-harness it is @bizos (setup).

⭐ If cross-harness saves you time, please star it on GitHub — stars help other developers find it and keep the BizOS team going.

What it does

• One message, many agents. @codex @claude review src/ starts both at once; one line per agent gives each its own task. @all, @auto and @compare:claude cover the rest.
• Real sessions, not screen scraping. Each agent runs through its vendor's own protocol — ACP, codex app-server or Claude Code's stream-json — so cross-harness sees every command, edit, plan step and permission request as data.
• Agents talk to each other. Every agent gets tools to ask another agent (with deadlock detection), ask you for a decision, share a project board and post its status. They work inside agent sandboxes because they go through files, not sockets.
• You approve in one place. Permission requests from all agents land in one queue. Choose allow once, always allow for this agent, or deny. Plan, ask, workspace and full modes decide what is automatic.
• Isolated by default. In git repositories each agent works in its own worktree on its own branch, so parallel agents never step on each other.
• Review, then merge. See an agent's diff, have a different agent review it, then merge with --no-ff. Dirty projects and conflicts are refused cleanly.
• Goals with evidence. Write down what "done" means: commands cross-harness runs, reviews by another agent, confirmations by you. cross-harness checks the work itself, sends failures back to the agent, and holds merges until the checks pass on the exact tree being merged.
• Anywhere. A terminal UI, a local web UI, cross-harness run for scripts and CI, and --ssh host to run every agent on another machine.

Review by another agent, then merge

Press d to see what an agent changed on its branch and m to merge it. cross-harness first asks a different agent to review the diff — here Codex reviews the BizOS CLI's fix — and merges with --no-ff only when you confirm.

The same, in your browser

cross-harness gui serves the same engine on 127.0.0.1: the agents and their live work, approval cards with one-click answers, diffs, reviews and merges, goals and the project board. It follows your light or dark theme and works in a phone-sized window.

How it works

One engine serves one project. Every interface talks to it through the same small client protocol, in the same process, over stdio or over SSH. Every agent is reached through the session protocol its vendor ships, normalised into the same seven events; agents reach cross-harness back through an MCP server (cross-harness mcp) that carries their tool calls as files, so it works inside their sandboxes. The protocol is specified in spec/ and the implementation is described in docs/architecture.md.

Install

Requirements: Node.js 20.11 or newer, git, and at least one coding agent installed and signed in (for example claude, codex, opencode, kimi).

Install the 0.2.0 preview from its release:

npm install -g https://github.com/AI-BizOS/cross-harness/releases/download/v0.2.0/bizos-ai-cross-harness-0.2.0.tgz
cross-harness detect           # shows which agents cross-harness found

Or from source:

git clone https://github.com/AI-BizOS/cross-harness.git
cd cross-harness
npm install -g .        # installs the `cross-harness` command; there are no dependencies

Update with cross-harness update (--check only reports a newer release); a source checkout is updated with git pull. The terminal UI and the web UI also mention a newer release, after checking GitHub at most once a day. The user guide covers installing, everyday use and troubleshooting step by step.

Quick start

cd your-project
cross-harness                  # terminal UI
cross-harness gui              # or: web UI in your browser

Then type a message that starts with the agents you want:

@codex fix the failing test in src/parser
@claude add edge-case tests for the parser; ask codex if the error format is unclear

Approve what they ask for, watch them work, press d to see an agent's changes and m to review and merge them.

Addressing agents

Mentions count only at the start of a line.