Skip to main content

Features

The feature that matters most is the one with no page of its own, because it is not a screen. It is a question.

Ask the graph

"Why is auth built this way?" "What did we decide about the payments retry policy?" "What happens when the alignment check can't run?"

One answer: what still stands, who decided it, and a link to where it was decided. Not seven Slack threads, a stale Confluence page and a Jira ticket that contradicts both.

You can ask from three places, and they read the same graph:

SurfaceHowBest for
An AI agent, over MCPsearch and friendsThe agent gets the decision before it writes code against a guess
The CLIalign search "<query>", align ask "<question>"You, in a terminal, mid-task
The web appSearch, filters, timeline, graph viewReading around a topic, onboarding, audits

No model of ours sits between you and the record. The MCP tools return the stored decision, not a generated summary of it. That is the claim worth relying on, and it holds unconditionally.

Retrieval is semantic where the query is a sentence, and keyword-based where it is a couple of words. Both paths break ties on a stable identifier, so the same query against an unchanged graph returns results in the same order - two decisions saved in the same bulk approval will not swap places between calls. That is a narrower promise than "deterministic": re-embedding the graph, or a change to the embedding model, legitimately changes semantic ranking.

The use case people recognise fastest

New joiner, week one. They can read the code to learn what was built. The graph is the only place that says why, and which of the three previous attempts is the one that survived. The same query pattern serves an incident review: what did we decide about this system, when, and what has superseded it since.

Get warned before you contradict something

Retrieval answers a question somebody asked. The Agent Alignment Layer answers the one nobody thought to ask: an agent (or a CI job) checks a proposed change against the graph before it lands.

This is the part general-purpose search cannot do. A contradiction is semantically opposite to the decision it violates, so the decision you are about to break is the one a similarity search is least likely to hand you unprompted.

And when the check cannot run, it returns unknown rather than quietly passing.

The rest, in the order you meet them

PageWhat it covers
CaptureGetting decisions in: @align, /align, reactions, meeting transcripts, and the per-platform matrix of what each vendor's API makes possible
DiscoverThe historical scan. Point Align at what you already have and review what it finds before anything is saved
SpacesScoping the graph to how your org is actually shaped, up to four levels, with cross-space dependency tracking
IntelligenceRelationships, conflict and supersession detection, tiers, health scores, impact and drift
SettingsConnectors, your own LLM provider and keys, members and roles, API tokens

What Align deliberately does not do

  • Reverse-engineer a decision from code. If nobody recorded why, the why is gone. Nobody can recover it. What Align guarantees is that from the day you connect it, the reasoning stops disappearing.
  • Write your code. An engineer still reads the decisions and interprets them. Align makes sure they are reading the right ones, and notices when a change contradicts one.
  • Listen passively. Every capture path starts with a human action or an explicit, per-resource opt-in that defaults to off.