Get started with Harness 1.5.0

What you can do

Attune Harness runs agent tasks, checks their results against criteria you supply, and keeps the output and check evidence in a receipt. Use it to plan work, build an accepted task, review a document against evidence, repair scoped files, or test a captured change. Inspect a saved task before continuing it.

Version 1.5.0, published October 19, 2026, includes merged Windows components for captured tests, file-memory readers, saved-memory storage and experimental native-memory transport. Full Windows parity remains deferred. Hosted Windows Server checks do not qualify Windows 11 desktop or PowerShell journeys, all storage/recovery cases, or live native-memory providers. Read the release notes for the component evidence and outstanding requirements.

Install the published release

Use Python 3.10 or later. Keep Harness and Attune AI in separate environments: they pin different MCP SDK versions. You do not need Attune AI for this example. Create a fresh environment instead of upgrading a retained evidence environment.

On macOS or Linux, in a new working directory:

python3 -m venv .venv
source .venv/bin/activate
python -m pip install 'attune-harness==1.5.0'

For a Windows installation, the equivalent PowerShell setup is:

py -3 -m venv .venv
.\.venv\Scripts\Activate.ps1
python -m pip install 'attune-harness==1.5.0'

The Windows commands are installation syntax, not a qualified PowerShell walkthrough. If activation is restricted, use .\.venv\Scripts\python.exe directly for the Python commands below.

In the environment you selected, check the version and run the local example:

python -c "from importlib.metadata import version; print(version('attune-harness'))"
python -m attune_harness
attune-harness --help

Expected result: The version is 1.5.0. The example returns a JSON receipt with status: verified, output 4 and an accepted arithmetic check. This deterministic example makes no model call; it establishes only that example's result. The published base install and these checks were run on macOS with Python 3.12.2 for this page.

If you use uv instead, install the isolated tool with uv tool install 'attune-harness==1.5.0', inspect uv tool list --show-paths, then run attune-harness --help. The Python commands above belong to the virtual-environment setup, not uv's separate tool environment. Installing the Python package does not install a Claude Code plugin or Codex skill; follow the host integration instructions for your chosen host.

Choose a workflow

Run a route's help in the same selected environment. The CLI guide describes configuration, checkpoints and recovery. Opening a form, saving answers, accepting intent and authorizing execution are separate actions. Browser forms do not dispatch builds. External and native participants require their applicable explicit grants; approving a plan does not authorize paid calls.

Read AI tools and integrations for a Typical use index, primary verbs, actual subcommands, ordered handoffs, sample prompts and expected evidence.

Choose retrieval deliberately

Voyage AI is the service Patrick uses and recommends. Its optional integration requires a Voyage API key and makes paid retrieval calls. Install 'attune-harness[voyage]==1.5.0' in the isolated environment only when you choose that integration, then follow the Voyage setup and configuration. Installation itself does not authorize execution or spending.

attune-rag is the explicitly selected local alternative. Harness 1.5.0 pins attune-rag==1.2.0 in its base dependencies; standalone attune-rag 1.2.1 is a separate release. Keep the Harness pin when following this page. Automatic fallback during a Voyage outage is planned; it is not available in 1.5.0. Select and configure the intended retrieval path explicitly.

Redis is optional. The all extra includes Redis and Voyage; experimental memory-native is separate. See the 1.5.0 dependency metadata for exact pins.

Continue with versioned guidance

• Published 1.5.0 release and PyPI package.
• 1.5.0 release notes: merged components, evidence and deferred Windows requirements.
• Preparation walkthroughs: retained candidate-wheel procedures and observed limits. These require their recorded artifacts and POSIX setup; they are not published-wheel first-use instructions. Use the installation above to start with published 1.5.0.
• Historical 1.3.0 task guides: earlier behavior, explicitly versioned and retained for existing users.
• Report a problem: include the installed version, route and redacted diagnostic. Keep credentials, private form links and project data out of reports.