Start with your way of working

Agent Harness is a harness of shared custom primitives for the two runtimes it supports today, Claude Code and Codex. Your personal stances are switches for behavior: communication, delegation, testing, autonomy and decision-making. The same selection resolves in both runtimes, and three of the nine axes — autonomy, delegation and cost — bind to enforcement rather than to prose.

Choose a runtime and inspect support

Claude Code and Codex are integration targets. Review compatibility before choosing a client; candidate projections are not proof of native support. Cursor and Grok are planned. You need your own runtime account and billing, Git and Python 3.9+. Native Windows is unsupported; WSL2 is unqualified. The harness does not sell or supply model access.

Install

One command clones the stable branch to ~/repos/agent-harness, writes a default configuration and previews the install without performing it:

curl -fsSL https://raw.githubusercontent.com/JakeSelby/agent-harness/stable/scripts/install.sh | sh

Read the preview it prints, then select which runtimes to manage and install:

cd ~/repos/agent-harness
bin/harness config set claude.manage true
bin/harness config set codex.manage true
bin/harness install --dry-run
bin/harness install

To clone by hand instead, or to see what each step of the script does, read installation ownership.

You can set either runtime to false; a Codex-only setup needs no Claude configuration. Install adds the selected missing runtime tools and synchronizes their configuration. On macOS it can also install Homebrew packages and VS Code; review install --help to skip those operations. When the tools are already installed, use sync --dry-run then sync instead.

Complete each selected client’s own sign-in. harness doctor reports files and configuration, not proof that a login or native hook is active. Start a new session and accept native hook trust where the client requires it. Use harness trust <repo> separately for the repository gate.

Make your first switch

bin/harness config set stances.voice answer-card
bin/harness stances --json
bin/harness sync

The same selected policy reaches both adapters. Restart the client to load changed global instructions. See resolved behavior before and after a switch, then author your own stance. Project/session overrides stay scoped to that invocation; syncing does not persist them into everyone else’s global files.

Use workflows and continue work

Shared workflows compose roles and skills. Claude exposes command projections; Codex exposes harness-<workflow> skill projections. Both use the same source. The handoff workflow writes shared task data; task continuation explains revision checks and verification.

Run harness diff for drift, harness usage for recorded measurements, and harness compatibility for qualification. Installation ownership covers custom configuration homes, adoption, uninstall and conflicts. Never replace an unmanaged file just to make sync quiet.

Your first session

From the repository you want to work in, start either installed runtime:

cd ~/some-folder
claude
# Or start Codex in that folder:
codex

Ask the agent to explain its effective stance choices, then try a small task. Research and planning can work without a code repository. Build and review workflows need repository context; opening a pull request also needs a remote and a signed-in GitHub account.

Then close the loop and ask which of your rules actually fired:

harness usage --rules

It prints one line per detector over the last 30 days — hits, the sessions that saw them and the share — with --by repo and --by stance regrouping the same hits by repository and by the preference variant in force. On a fresh home it says no measured sessions in the last 30 day(s) until a session or two has been recorded; --rescan backfills from transcripts you already have. Usage explains the annotations and what the figures do not yet support.

What it costs

Model access is billed by your provider or covered by your subscription; rate-limit windows and context usage still matter. The harness supplies policy, not an AI model or paid access. Usage shows observed measurements without inventing missing values. The preferences guide explains the cost stance. How it works and sandboxing explain the controls. harness uninstall restores owned settings where possible and reports conflicts rather than overwriting your edits.