Dots Studio

Four interactive workbenches. One input you can inspect and replay.

imageScenario Lab using a CSV

Try the panels in a minute

Node 22 or newer is required. No API key is needed.

git clone https://github.com/aichance/business-ai-recipes.git
cd business-ai-recipes/tools/dots-studio
npm ci --ignore-scripts --no-audit --no-fund
npm run build
npm test
npm run preview

Open Scenario Lab, Evidence Canvas, or Run Lens, or Proof Pack Builder. Choose Open your file, select items, then Export full-input replay. Save the file or copy its visible JSON if the host blocks file downloads. Change one assumption and compare results. Stop the local server with Ctrl-C.

The full-input replay contains all supplied input, including unselected items, to reproduce the full calculation. Share selected context sends only checked items (plus linked sources and next checks) to a connected host. It is disabled in the local preview. Neither operation automatically posts a message, executes a repair, or changes an external system.

Replay without the UI

From the repository root, after installing the dependencies above:

node tools/dots-studio/replay.mjs tools/dots-studio/fixtures/own-sales.scenario

This pack came from the UI using the second synthetic CSV: baseline gross margin 400.00, price assumption +1%, new gross margin 405.60, difference 5.60. Only Workshop is selected for context. Guide remains in the full-input pack so its contribution to the total can be recomputed.

The CLI validates the input and recomputes from it; it does not trust a saved result or execute a command from the JSON. Open the pack in its matching panel to restore assumptions, selection, decisions and replay position.

Plugin source and extension surfaces

plugin.json + portable mcp.json describe a local stdio MCP process. Prepare dependencies and build assets in this directory before adding it to a host. A host must support local processes, MCP Apps and OpenAI Extensions. Use that host's documented local-plugin installation flow with this directory; source publication does not install it into an account or provide a remote endpoint.

The four open_* tools register sidebar/global, thread-panel and file entrypoints (.csv/.scenario, .evidence, .dotrun, .proof). The UI implements the granted file-resource bridge and selected model-context update. The four analyze_* tools also open the matching panel; replay_pack recomputes a pack. HTML resources are bundled with no remote script or font fetches.

Verified: build, fourteen tests, discovery of nine tools and four HTML resources, local preview import, condition changes, selections and replay JSON. Pending: installation and a complete run in an actual dots account/host, host file permissions and context delivery. The preview displays HOST UNVERIFIED for this reason.

Dots can use supported plugins available to its computer/account, but that does not guarantee that a particular local stdio plugin works in its cloud computer. Official dots guidance, Extensions specification, SDK.

Bring a different input

• CSV: exact header name,unit_price,unit_cost,units; 1–1000 rows; nonnegative amounts with up to two decimal places and integer quantities. All rows use one currency/unit. Assumptions range from -80% to +100%. Prices round per unit to cents; quantities round to integers.
• Evidence JSON: title, claims: [{id,text,sourceIds}], sources: [{id,title,url,excerpt,checkedAt?}]. Only HTTP(S) links are accepted; links and excerpts are supplied, not fetched. linked does not mean true. Decisions are recorded as the user's choice, not an AI approval.
• Run JSON: title, steps: [{id,title,dependsOn}], events: [{step,at,type,note?,artifact?}]. at is nonnegative seconds; type is start/success/failure. The whole log is checked before prefix replay; dependencies must finish successfully before their dependents start. Artifact names are declarations; files are not opened or verified.
• Proof JSON: title, input, before, after metric arrays with {id,title,value}, and checks: [{id,title,status,note,metricIds,artifact?}]. Status is supplied evidence; unverified remains unverified and missing before/after values are shown instead of being inferred.

Files are capped at 200 KB. Large selected context is rejected with a smaller selection instruction; it is never silently truncated. The tool has no live dot log collector, background notification subscription, truth checker, forecasting model or auto-repair executor. It works on explicitly supplied input and leaves decisions with the user.

Why this instead of pasting everything into a chat?

The panel makes assumptions and the selected scope visible. A spreadsheet can do the same arithmetic, and a text editor can read the same log; this prototype adds an inspectable UI, dependency-aware replay and a portable handoff pack. It makes no measured time-saving or adoption claim.

50 ranked extension ideas are the next experiments; only the four workbenches above are implemented. AI-assisted work by Naoya / jokv213.

日本語: CSVの条件を動かす、主張と根拠を選ぶ、失敗したrunを時間順に追う、証拠パックを比較する4画面。 結果だけでなく入力・条件・選択をJSONで持ち帰り、同じ計算を再現できます。 現版はローカル画面とMCP接続まで検証済みで、実際のdots内の利用は未検証です。