Propose · planned
The agent drafts a decision with its reason and source. It stays a draft until its owner confirms it.
For designers and developers using AI coding assistants, a choice made in one chat can get lost in the next. Airouz is exploring a shared log of human-approved decisions that agents can consult before they build.
Product concept in development. The workflow and interface below illustrate the intended product; they are not a live application or a claim of available integrations.
claude-code/7f3aThe goal: let a person approve a decision once, then make it available to other agent sessions. The steps below describe the proposed workflow.
Planned flow: an agent proposes a decision, a person approves it, and other sessions consult the saved Markdown before checking their work.
We are designing three MCP tools: propose, context and check. MCP is the connection through which an AI assistant could read approved decisions and return possible conflicts for human review.
The agent drafts a decision with its reason and source. It stays a draft until its owner confirms it.
Before a task, a session loads only the confirmed decisions that touch what it’s about to change.
After a task, output is compared with confirmed decisions. A contradiction comes back with the decision and who owns it.
The proposed storage format is Markdown in your repository, so decisions could be reviewed and versioned alongside code. This fictional example illustrates the format under consideration.
---id: D-014what agents citestatus: confirmedonly a person sets thisowner: designwho to askconfirmed: 2026-10-07T10:42source: claude-code/7f3athe session it came fromexpires: 2026-12-31retires instead of going stale--- Sign-in uses magic links. Do not add a password field.Why: fewer reset tickets; one less field on mobile.The design goal is to preserve deliberate choices, their reasons and their owners, so a future session can distinguish an approved decision from a suggestion.
Airouz is an early-stage, self-funded project started around June 2026. It is not incorporated. The landing page presents the product direction; public product access and integration validation are still ahead.
Checks are designed to run inside each person’s own agent session, and decisions live in your repo. Before anyone installs it, we’ll publish exactly what leaves your machine.
Claude Code and Cursor are the first intended clients. The planned Claude workflow is to propose a decision, ask a person to approve it, retrieve relevant Markdown decisions through MCP, and check task output for possible contradictions. Compatibility and reliability still need to be validated; this is not an announced live integration.
The intended rule is that an agent can propose a draft, but a person must approve it before other sessions treat it as a decision. This approval mechanism is part of the planned implementation.
We are exploring Slack or email approval for teammates who do not use a coding assistant. These integrations are planned, not currently available.
The proposed format includes an owner and an optional expiry date. Renewal, retirement and reminders remain design goals.
Pricing and public availability have not been announced. Contact us to discuss the problem and hear about future access.
We’re looking for a few small teams where designers and developers build with agents together. Tell us where your decisions drift.