Harness docs and project references
Start with Attune Harness. Find guidance for useful work, execution permissions, checkable results and retained context. Earlier Attune AI references follow for existing users.
Start with Attune Harness
Define a useful next step, build or repair within scope, check the result, and return with the context intact. Choose a workflow from the guide; command syntax is available when you need it.
Claude Code and Codex have documented setup paths. Native planning and building remain experimental; Antigravity integration is unverified here. Intent acceptance is separate from paid execution permission.
Harness does not yet replace every Attune AI workflow. Review the migration boundaries before changing an existing setup.
Harness Quick Start
Install Harness and try a small local check before connecting an agent. This setup uses the released 1.4.0 package.
You need Python 3.10 or later. Start in a new working directory and keep Harness in its own environment, separate from Attune AI. The commands below are for macOS and Linux.
1. Create an environment
python3 -m venv .venv source .venv/bin/activateWindows PowerShell
With Python 3.10 or later installed, use these commands instead:
py -3 -m venv .venv .venv\Scripts\Activate.ps12. Install and confirm the version
python -m pip install 'attune-harness[all]==1.4.0' python -c "from importlib.metadata import version; print(version('attune-harness'))"The version check should print
1.4.0. This recommended install adds Redis and Voyage support. The local example below needs no API key. Redis use needs a running Redis service; Voyage retrieval needs a Voyage API key and makes paid calls. Integration requirements are covered in the released installation guide.3. Run the local example
python -m attune_harnessExpect a JSON receipt with a
verifiedresult. A deterministic worker supplies the answer and a separate arithmetic check assesses it. This example makes no model call and needs no API key; it demonstrates the check and receipt, rather than measuring an AI agent's reliability.
Continue with your agent
Choose the Claude Code or Codex setup in the release guide. Installing the Python package does not install an agent plugin or skill. Develop the specification and agree with the scope before authorizing execution.
Standalone help reader
Attune Help reads existing .help/ templates. If you need the reader independently, see the standalone package reference.
Attune Harness FAQ
What is Attune Harness?
Attune Harness is an open-source toolkit for AI-assisted engineering. It helps you define scoped work, record permitted actions, and inspect the results and the checks behind them.
Can I use it for specification development?
Work with your agent to develop requirements, constraints and acceptance criteria, then agree with the scope before authorizing execution. Harness supports planning, building, repair and test journeys. Native planning and building are experimental in 1.4.0; use the documented agent paths and review their limits.
How do I get started with Harness?
Follow the Harness Quick Start: create a separate Python 3.10+ environment, install attune-harness[all]==1.4.0, confirm the version and run the local example. Redis use needs a running service; Voyage retrieval needs an API key and makes paid calls.
What does a receipt tell me?
A receipt records the task outcome, captured output and check evidence. Look for independent checks alongside the output, and keep unchecked behavior separate. A passing check establishes only what that check assessed; it does not guarantee a correct change or a reliable model.
Which agents and models can I use with Attune Harness?
Harness has documented setup paths for Claude Code and Codex. Models and authentication depend on the host and participants configured for the task. Installing the Python package does not install agent plugins or skills. Native planning and building are experimental in 1.4.0. Browser forms in that release support draft intake, preview and intent approval; browser build dispatch is not available. Intent approval is separate from permission to execute paid work.
Does Harness replace Attune AI?
Harness runs without Attune AI, but does not yet replace every Attune AI workflow. Keep the two packages in separate environments and check the migration guide before changing an existing setup. The earlier Attune AI references remain available for existing users.
Does Harness keep all my data local?
Harness can read local memory and keep receipts on disk. Redis may run locally or remotely, depending on your setup. Authorized model calls or Voyage retrieval can send selected task context, source text or queries to a provider. Review the chosen workflow, destination and permissions before execution.
Earlier Attune AI references
The questions below describe earlier Attune AI tooling and companion packages. Their workflows and commands apply to those projects.
What is attune-ai?
Attune AI combines specification development, AI workflows, project memory, retrieval and verification tools. It is a separate project from Attune Harness. Use the retained references for your installed version, and inspect generated output and the checks that ran.
How can I check generated content for drift?
Retrieval through attune-rag supplies source context for review. Source hashes can flag tracked templates for maintenance when the corresponding code changes. These checks cover particular sources and properties; they do not establish the accuracy of every generated answer.
Do I need attune-ai to read templates?
No. The standalone attune-help reader can read existing help templates without the full Attune AI framework or an API key. Point it at the template directory described in its package reference. Generating or updating templates is a separate task.
Can I write templates by hand?
Yes. Use maintenance: manual in template frontmatter to mark a page as human-maintained and have regeneration skip it. Review the maintenance instructions for your installed Attune AI version before updating templates.
How does staleness detection work?
Generated templates can record source hashes. Maintenance compares the recorded and current hashes to identify affected templates. A stale flag identifies a review or regeneration task; it is not proof that the replacement content is correct.
Is it free? What do I need to run it?
Attune AI is open source under Apache 2.0, subject to the license conditions. Runtime requirements depend on the workflow; model providers and other external services may require separate credentials and incur charges.
What Claude Code skills are included?
The retained Attune AI skill reference covers development tasks such as review, testing, specifications and documentation. Skill availability depends on the installed plugin version and host. Inspect that installation rather than treating a website count as an environment check.
Where do I install attune-help from?
The standalone attune-help reader is available on PyPI. See its package reference for installation and template-directory setup. Installing a Python reader is separate from installing a host plugin.
What happened to attune-author?
Its AI authoring capabilities were consolidated into attune-ai in July 2026, and the standalone package is archived. Already-released versions stay installable from PyPI, but no further releases are planned — for template authoring, install attune-ai and use the author-feature skill or /coach maintain.
Choose a Harness workflow
Start with a concrete goal and a result you can assess. The guide explains the available workflows and their limits.