Skip to content
StoryShelf

AI insights

View as Markdown

StoryShelf can read a build’s results and tell you what it thinks: is this change intended, does it need a human look, or does it look like a regression? It can also summarize a project’s health over the last 30 days. Both are advisory only — they never approve, reject, waive a failing play function, or override a changed baseline. You stay in charge of every merge decision.

The review page with an AI triage panel above the snapshot list: verdict, summary and per-snapshot notes The review page with an AI triage panel above the snapshot list: verdict, summary and per-snapshot notes
AI triage on the review page: a verdict, a short summary and notes per snapshot.

What you get

Section titled “What you get”

Build triage

A verdict (likely-intended, needs-review, likely-regression), a short summary, and a note per snapshot — right above the review workspace. How it works

Project health

A 30-day digest with a verdict, a 0–100 score and trends, on the Library and Builds pages and as a badge on the project card. How it works

Your models, your keys

OpenAI, Anthropic, Google, Bedrock, Vertex, Azure, a gateway, or a local model. StoryShelf never sees your keys. Set it up

MCP server for agents

Let Claude Code, Cursor and other agents read builds, diffs and triage, and comment on reviews. Read-only plus comments; approving stays human. Connect an agent

Private by design

Off until a site admin enables a profile per project. Secrets are redacted, output is plain text, and traces never contain prompts or screenshots. Privacy and safety

Turn it on

Section titled “Turn it on”

  1. Give the server a model. Install @storyshelf/ai and a provider package, create an AI surface, and pass it to createShelfApp. See Setup and providers.

  2. Choose a profile for a project. A site admin opens Project settings → AI (or sets aiProfile over the API). Until then the project sends nothing anywhere.

  3. Review as usual. The AI triage panel appears on the review page, and approvers and admins can generate or regenerate it. Generate the health digest from the project’s Library page.

Project settings, AI tab: a select to choose the AI profile for the project Project settings, AI tab: a select to choose the AI profile for the project
Site admins choose which profile (and therefore which models) a project uses.

Cost and control

Section titled “Cost and control”

Everything is bounded: a daily token budget across all instances, a per-project hourly call limit, capped evidence size, and identical evidence is served from cache for free. Site admins get alerts at 50%, 75%, 90% and 100% of the budget and a usage page. See Budget and usage.

Where to go next

Section titled “Where to go next”
Setup and providers createAi, profiles, models, provider options and credentials.
MCP server Connect AI agents to builds, diffs and triage over stdio or HTTP.
API Endpoints, roles and the guard order for scripts and CI.
Free AI for testing Try it with a free-tier hosted model or a local one.