Skip to content

Principles

ALIGN is a decision ledger for teams that build software with AI agents. These promises were written before the first customer, and the product is built so that breaking one would take a visible change to its code, not a quiet change of policy.

Six promises

  1. Neutral. It works with Claude Code, Cursor, Codex and any MCP agent, and never ships an agent of its own. The memory your harness keeps is one of the sources it reads.
  2. Receipts. Every entry carries an exact quote at an exact location: transcript lines, a commit, a pull request, a Slack or Linear message, a recorded event or a confirmation. The database refuses an entry without one.
  3. Verified. Confirmed decisions get deterministic checks where possible ("ioredis is not imported under lib/cart/"), run on every push to the default branch and every pull request. A decision the code no longer honours becomes a conflict with receipts on both sides; one that cannot be checked is marked unverified.
  4. Yours. Export the whole ledger as Markdown or JSON at any time, from the console or with align export. Delete a session, or the organisation, yourself.
  5. Permanent ids. Decision ids such as AL-12 live in your commit trailers and outlive any tool, including this one.
  6. Quiet. One check and at most one comment per pull request, edited in place. Notices are budgeted. No per-developer productivity metrics in the product, analytics, logs or exports, ever.

Models propose, people confirm

A model may suggest that a session contained a decision. Code checks that the quote it cites really is in the source, a second model can veto it, and the proposal waits. Only a person confirms, rejects or supersedes a decision: with one key press in align confirm, a checkbox in the pull request comment, a click in the console or a reply in Slack.

What it will not be

  • Not an agent and not a harness. It never edits code or runs tasks.
  • Not a tracker. Linear and GitHub Issues stay where intent lives.
  • Not a code reviewer. The pull request check asks one question: does this change honour or contradict a recorded decision, and did it make a new one?
  • Not people analytics. Seat counts for billing are aggregates and appear nowhere per person.
  • Not screen recording, keystroke capture or file upload. Capture never sends your files' contents: tool calls become placeholders and fenced code blocks a line count. What it does send, your prompts and any text pasted without a code fence (after secret redaction) included, is in what is captured.

Read how your data is handled and the privacy policy.