dumpr

Type a brain dump → Claude parses it into Notion rows → review/edit → write.

Built on the Claude Agent SDK. Uses your Claude Code OAuth, so there's no API bill — inference runs on your Max subscription.

Why

every productivity system fails at the same point: capture and triage live in the same moment.

you have a thought, and the system asks you to also pick a category, priority, due date, energy level, and project. so you pick none. the thought rots in an untagged row.

i've been a notion guy for years and watched my own database become a graveyard of half-captured ideas. the structure was there, i just never had the energy to fill it in mid-flow.

dumpr fixes that. dump first, let claude do the boring part, review the table, ship to notion. capture and triage stay separate, the way they should be.

Setup

1. Notion database

Create a database with these 7 properties — exact names, case-sensitive:

Property	Type	Options
Title	Title	—
Type	Select	Task, Project
Domain	Select	(you choose — see step 3)
Priority Level	Select	P1 – Critical, P2 – Important, P3 – Normal, P4 – Low ⚠
Effort	Select	Low, Medium, High
Status	Select	Backlog, Planned, In Progress, Blocked, Done, Dropped
Due Date	Date	—

⚠ Priority values use an em-dash (–), not a hyphen (-). Copy-paste the exact strings from the table above. Notion select options match exactly, so a hyphen will silently fail with a write error like Priority "P1 - Critical" is not a valid option.

2. Notion integration

1. Create an internal integration at https://www.notion.so/profile/integrations with capability Insert content.
2. Open your database → … menu → Connections → grant your integration access.
3. Copy the integration token (secret_…).
4. Find your data source ID (this is what the v5 SDK writes to — not the database ID). Open the database as a full page, then … menu → Copy link to data source, and pull the UUID from the copied URL.

3. Customize your domains (the personal axis)

Domains are the life-buckets that mean something to you. The repo ships with 14 generic example domains (Work, Clubs, Community, Coursework, Family, Health, Deen, …) as a starting point — your real life buckets are probably different. Edit three places:

1. lib/types.ts — DOMAIN_VALUES array (the canonical list).
2. lib/notion.ts — DOMAIN_ICONS map (one emoji per domain; sets the per-page icon in Notion).
3. skills/SKILL.md — the "Domain inference" section near the bottom. This is what the AI uses to decide which domain a dump line belongs to. The most important of the three.

For each domain you set, also add it as a select option on the Domain property in your Notion database, otherwise writes will fail.

4. Environment

Copy .env.local.example → .env.local and fill in your values:

NOTION_TOKEN=secret_<your token>
NOTION_DATA_SOURCE_ID=<your data source UUID>

5. Claude Code OAuth

This app uses your Claude Code OAuth — there's no ANTHROPIC_API_KEY, inference runs on your Max subscription. If you don't already use Claude Code:

npm install -g @anthropic-ai/claude-code
claude /login

This creates ~/.claude/config, which the agent SDK reads automatically.

6. Run

npm install
npm run build
npm run start

Open http://localhost:3000 and bookmark it. For development: npm run dev.

How it works

lib/skill.ts reads skills/SKILL.md, appends a JSON-output override, and sends both as a system prompt to Claude via @anthropic-ai/claude-agent-sdk. The model returns a JSON array of typed items. Zod validates the shape, the UI shows a preview table for inline editing, and on confirm the rows are written to Notion sequentially via @notionhq/client.

Tests

npm test

34 tests covering SKILL.md prompt assembly, Item → Notion property mapping, batch writes with stop-on-first-failure, agent-SDK orchestration with JSON validation, and both API routes. All tests mock the agent SDK and Notion client — no live services touched.

License

MIT. See LICENSE.