What you can set up with SubKit.
Use cases, not features. Each row describes a specific situation and the workflow you set up for it: Example to set up means you set up the workflow yourselves – with your own access and approval. Multiple systems means the workflow connects multiple systems. Future direction means this part is the direction in which SubKit is being developed and has not yet been fully released to the public.
All cases are possible workflows to set up – not ready-made templates included with the product, customer references, or claims about any particular operation.
- 01Personal assistantOne person, their own tools
- 02The company that answers questionsAnswers with sources – or an honest “not documented anywhere”
- 03Workflows across multiple systemsAn agent builds it, an agent can start it – the path is defined
- 04Domain-specific AI for an entire area of workShared context so every workflow within it works as intended
- 05Connect a knowledge baseDocuments become knowledge that AI clients can build on according to their permissions
Personal assistant
The smallest useful SubKit setup: one person pointing it at the tools they already use. Even on your own, you do not have to explain your knowledge and way of working again for every task. Approval is still part of the process – it defines what the workflow may access.
| Situation | What SubKit does | Systems | Classification |
|---|---|---|---|
| A pricing page has no API, and the price changes | Opens the page in a headless browser, reads the price, and emails you when it changes | Browserless + Gmail | Example to set up |
| A provider's receipts are scattered across months of email | Filters messages with attachments into a list | Gmail | Example to set up |
| PDF invoices are scattered throughout the inbox | Searches for emails with PDF invoices every evening, names each file by date and sender, and saves it in a dedicated Drive folder | Gmail + Google Drive | Multiple systems |
| Know on Friday what ran during the week | Counts the week's runs, lists failures, and emails the digest | Gmail | Example to set up |
| “What is on my plate today?” | Retrieves your open tickets with all details | Jira | Example to set up |
| A long article, five minutes to spare | Opens the page and has it summarized in a defined AI step | Browserless + Anthropic | Example to set up |
| Infrastructure costs grow unnoticed | Emails a daily overview of resources, grouped by project | DigitalOcean + Gmail | Example to set up |
| A dashboard should be in your inbox every morning | Reads the page at 07:00, extracts the values, and emails the summary | Browserless + SendGrid | Multiple systems |
The pattern: describe it once in a sentence, review and approve what SubKit has built, and from then on it runs – on request, on a schedule, or when something arrives.
The company that answers questions
Your team asks questions where it already works – Slack, Teams, Google Chat, an AI client in the browser – and receives answers from your own confirmed knowledge. Each person sees only what they are allowed to see.
| Situation | What SubKit does |
|---|---|
| “How do we usually do this here?” | Searches your sources and answers with a source citationthat states what was read – or clearly says the answer is not there instead of inventing one. |
| The team lives in chat | The SubKit bot answers questions and starts saved workflows right there. By design: A chat window executes things; it does not rewrite them. Changing a workflow goes through something a person can read and review. |
| “Who saw what?” | The same permissions and history as at every other entry point. Chat is not a back door. |
What makes this shared context rather than a chatbot: The knowledge is curated (a person confirms what is accepted), answers cite their origins, and every interaction is permission-checked and recorded.
Workflows across multiple systems
What a defined workflow can still do:
| Situation | What SubKit does | Classification |
|---|---|---|
| A change may be applied only after human approval | Opens the proposed change, waits up to 24 hours for an approval signal, then applies it – or aborts | Example to set up |
| A new customer needs to be enriched and reported | Stripe customer → HubSpot contact → Slack message, one workflow across three systems | Multiple systems |
| The supplier's price list arrives every week, but the shop lags behind | Reads the sheet, compares it with the catalog, updates prices, hides sold-out items, and reports discrepancies above the threshold instead of writing them | Multiple systems |
| The weekly report is manual work on Monday morning | The previous week's visitors, orders, and revenue from analytics, the shop, and payments, in one email at eight | Multiple systems |
| A website inquiry should reach sales within seconds | Signed form notification → CRM contact, if new → message in the sales channel → acknowledgment | Multiple systems |
| Access should be available immediately after a paid checkout | Verifies the Stripe signature, confirms payment, enables access in your own system, and sends the welcome email | Multiple systems |
| Notifications from any system, but only the important ones | Receives signed notifications via webhook, keeps those at or above the chosen level, and posts them to a Slack channel | Example to set up |
| Trial users need contact on days 1, 3, 7, and 14 | A daily run with saved state per contact to control the sequence | Example to set up |
| A transaction must never be half-completed | Debit and credit in a single database transaction | Example to set up |
| A failed payment needs immediate attention | Signature-verified notification from the payment provider → Slack alert without personal data | Example to set up |
| Market signals should become published content | A chain of eight workflows: collect → normalize → analyze → generate → publish → acceptance testing | Example to set up |
| Metrics are in five systems, and the report is manual work | Daily, weekly, monthly, and annual summaries across analytics, search, repositories, and infrastructure – plus cleanup of old states | Multiple systems |
| Contacts are duplicated or outdated | Detects duplicates and outdated entries, presents a list for review, and cleans up only after confirmation | Example to set up |
What this is built with: Calls to each system with consistent retries and logging; loops and transformations for lists; fixed branches (“if refund, otherwise charge”); parallel calls; workflows that call other workflows; a step that waits for a person and aborts after a deadline; validation rules that fail a run rather than pass on incorrect data; and filters that mask personal data at the tenant boundary.
Domain-specific AI for an entire area of work
A single workflow answers one question. An area of work – your expense rules, your editorial style, your review framework – provides the shared context that makes every workflow within it work correctly. You describe it with three things: a Connector (the system that holds the source of truth), Skills (rules the AI follows), and Sources (documents it looks things up in).
| area of work | The skill specifies | The source provides | Classification |
|---|---|---|---|
| Expense approval | Two approvals above €1,000, receipts required | The expense policy as a PDF – limits, daily allowances | Example to set up |
| product catalog | Unit prices for quantity-based products; crossed-out prices according to guidelines your team has reviewed; price changes over 15 percent are reported, not written | The supplier's price list as a sheet, product data from the shop | House rule, to be reviewed by the team |
| Accounting | The mandatory invoice information as specified by your accounting team; handling e-invoices according to your requirements | Receipts from the inbox, the export from Lexware Office or sevDesk | House rule, to be reviewed by the team |
| Newsletter | Advertising only with consent, confirmed subscription before the first send, an unsubscribe link in every email – as reviewed by your team | The customer list from the shop, the segment in the newsletter tool | House rule, to be reviewed by the team |
| Editorial house style | Punctuation and style rules the AI must follow | The body of text it can quote | Example to set up |
| Provider-specific quirks | “Stripe reports validation errors with status 200 and an error object” – foundational skills like these for each provider | Provider documentation, ingested per connector | Example to set up |
To be clear: You bring your own domain expertise, and the trio above is how you do it. House rules relating to legislation are your team's requirements – they do not replace legal advice, and SubKit does not guarantee their legal correctness.
Connect a knowledge base
Bring knowledge in, keep it up to date, and allow AI clients to build on it according to their permissions.
| Situation | What SubKit does | Classification |
|---|---|---|
| The knowledge is on a website | Reads the sitemap and converts every article into clean Markdown for ingestion | Example to set up |
| A document is available only as a rendered web page | Converts it to Markdown for further processing | Example to set up |
| Every session with a coding agent starts from scratch | Claude Code or Codex sessions start with a summary of your confirmed rules and can look up instructions instead of inventing them | Future direction |
| What a session learned disappears when it ends | Collect an observation → review and prepare it in chat → a person explicitly confirms it → accepted knowledge or a skill available to the next session. Raw excerpts are not confirmed knowledge. | Future direction |
| Nobody knows where the knowledge gaps are | A weekly coverage report on the knowledge base | Example to set up |
| A source has changed, but nobody knows when or by whom | Every source has versions with authors; any version can be restored as a new one, and an address is reread at the configured interval | Built in |
| A model should read the documentation itself | The engine serves its documentation in a form language models can read directly | Built in |
The loop that develops: Knowledge underpins design, runs reveal patterns, and people confirm what has proved useful. The complete process from collection to knowledge is the future direction and has not yet been publicly released.
Start with a specific task.
Back to the overview
How SubKit works, what it does not do, and why the approach holds up.
Back to home →