Open-source tools for AI-assisted engineering

Turn a goal into a plan.
Put agents to work.
Check the results.

Plan a change, review a document, repair files, run tests or return to saved work with your assistant in Codex or Claude Code. Harness keeps the output and check results so you can see what happened and decide what to do next.

Choose your next task

Get started Work with us

The lead project · Attune Harness

Make progress you can review.

Choose a useful goal, agree on the work and checks, and review the result.

Start with the problem you want to solve. Your assistant helps prepare the work, presents what needs your review and uses Harness to carry out the steps you authorize.

Turn an idea into a plan.

Develop the goal, affected files, tasks and checks with your assistant before implementation.

Build in reviewable steps.

Give configured agents the accepted tasks, then inspect their work against the agreed checks.

Check a document’s claims.

Compare a document with your criteria and supporting sources. Inspect findings and unresolved questions.

Repair a problem and test the fix.

Check a proposed repair against a test fixed before the repair begins.

See which tests ran.

Preview the tests selected for a code change, then inspect their output and results.

Pick up saved work.

Read the saved decisions and results before choosing the next action with your assistant.

See what passed, what failed and what needs another look. Harness keeps the output and check results together so you can review the work before deciding what to do next.

Decide what runs.

Accepting intent and authorizing execution are separate decisions. Paid calls need their own authorization.

Understand the checks.

A passing check establishes what it tested. It does not guarantee every output is correct.

Published: Attune Harness 1.6.0. Get started · Published package

Harness 1.7.0 is in preparation. Use the published release for installation. Explore the 1.7.0 preparation guide.

Browser forms guide intake and intent approval. Each workflow retains its own execution permissions and checks.

Read the evidence and limits

In one exporter comparison, command-line checks found three defects that direct serializer checks had missed.

The result applies to that example. General model reliability needs separate assessment.

Read the comparison and its limits

A passing check establishes what it tested. Native planning and building remain experimental. Claude Code and Codex have documented setup paths; Antigravity integration is unverified in this review. Native MCP panel delivery remains unverified. Review the qualification guide.

Context and useful next steps

Find the next opportunity in work you already have.

A bounded workflow with your agent: bring relevant context, review possible follow-up work, and decide what is worth pursuing.

Bring the relevant context.

Use project evidence and supported memory sources you explicitly provide or configure.

Keep the next step visible.

Retain an opportunity-review checkpoint with its goal, progress, evidence and unfinished review.

Choose what to pursue.

Assess the proposed benefit and uncertainty. A saved suggestion grants no execution authority.

Scoped recall and saved review checkpoints are supported. Automatic access to all chat history is not implied; the richer opportunity-ranking and selection journey remains proposed. Review the current boundary.

Help shape what comes next

Three ways to take part.

Proposed paths for contributors, potential pilot customers and open-source sponsors.

Start by building together

Collaborate

Help improve specification development, planning, build and repair workflows, and independent checks. Bring a reproducible problem, a design question or a contribution.

Explore collaboration

A starting conversation could identify a useful contribution, its scope and how to check it. Documentation, usability and real workflow evidence matter alongside code.

Start a conversation

Explore a possible customer engagement

Discuss a pilot

Explore one workflow guided by a specification, from a clear goal and agreed scope to evidence you can review. Define what a possible pilot would need to demonstrate.

Explore a pilot

Possible starting points include developing a specification with explicit acceptance criteria, reviewing a generated document against project evidence, or checking a scoped agent-assisted change. Scope, responsibilities and commercial terms would need separate agreement.

Start a conversation

Supporting tools and earlier work

Project references.

Harness is the current focus. Supporting libraries and earlier project references remain available.

Earlier project · Memory and workflows

Attune AI

v16.4.0

Reference for the earlier project’s memory and workflow tools. Existing code and documentation remain available.

Attune AI reference

Structured questions and decisions

Attune Forms

Make questions and choices easier to review, with structured answers validated on return.

Explore Attune Forms

Retrieval with citations

Attune RAG

Retrieve relevant project sources and return citation records for the next step.

Explore Attune RAG

Built by Patrick Roebuck

Developed in the open.

The direction is practical: make agent work easier to authorize, inspect and evaluate. The next step is to find useful problems to work on together.