Open source · No cost during beta
The Last SaaS
A Headless SaaS for Agents and Assistants.
Keep records, files, users, permissions, notifications, and audit logs in one system that any AI assistant can manage. Describe what you need, and the assistant handles both the initial setup and ongoing changes, from adding a field to inviting a new user.
Questions go to the assistant, which reads the records and answers with a list, a chart, or an answer combining Last SaaS data with other sources, depending on the question. The assistant is whichever one is already in use.
Here's what setup looks like. Expand a row for the details.
Setting that up now.
Created collection · Gear
- Fields
- Item · Serial · Status · Checked out by · Purchase price
- Status
- On the shelf → Checked out → In repair
Added rule · Team ↔ Gear
- Can
- view and check out any item
- Except
- the Purchase price field
Invited 3 teammates
- As
- members
- Status
- invitations sent
Scheduled digest · Mondays 9:00
- Sends
- what's checked out, and by whom
- To
- you, by email
Done. Gear tracking is live. Your teammates have invitations waiting; they can check anything out and never see what it cost. Your first digest arrives Monday at 9. Tell me what's on the shelf whenever you're ready, or forward me the receipts.
Added. It's on every item now, blank until you fill it.
Updated collection · Gear
- Added
- Last serviced
- "Keep the club's waivers and who has signed."
- "Plan the launch: tasks, owners, a Friday status ping."
The Last SaaS is a server that holds records and files, tracks who is on the team and what each person can see, sends notifications, and logs the changes. It does not provide a user interface on its own; an assistant connects over MCP and operates it.
- records
- files
- people
- permissions
- notifications
- audit logs
01The assistant as the frontend.
Most SaaS is built from the same parts: data, files, users, permissions, and notifications, with a user interface on top. That interface is a large part of what the subscription pays for, and it is work a model does well. Asked how the quarter is going, the assistant reads the records and draws the chart, from the data as it stands at that moment.
Read collection · Invoices
- Query
- issued Jun 1 – Aug 31
- Fields
- Amount · Status · Client
- Matched
- 12 records
You've collected $18,400 of the $23,900 invoiced. Two invoices are overdue, both from Meridian.
Assistants connected over MCP read the same records and can draw the same chart.
02The pieces these tools are made of.
The system provides six primitives:
- Records. Structured lists of anything from gear to invoices. The assistant sets up the fields for whatever is being tracked and keeps the entries valid.
- Files. Receipts and contracts, attached to the records they belong to.
- People. Teammates join by invitation and work from the same records, each with their own assistant or none at all.
- Permissions. Rules apply down to the row and the field.
- Notifications. Due-date reminders and weekly digests, recurring or one-off.
- Audit Logs. Changes are recorded with who made them and when, whether it was a person or an assistant.
This is also the difference between The Last SaaS and a spreadsheet or a Notion base with AI attached. Those are apps designed around a UI, with an assistant added afterward. This is a server designed for agents from the start, and the user management, permissions, notifications, and audit logs make it a complete system.
03Each person gets their own view.
Permissions apply to people and to their assistants. A bookkeeper's assistant can read the invoices and cannot open payroll. The rule is stated once and enforced on each request.
- Rule · The bookkeeper
- can view and add records in Invoices
- can attach files to any invoice
- not the Payroll collection
- not the Margin field on invoices
A rule can be applied as far down as a single field on a single record, and it covers any assistant acting for that person. History records the changes and who made them.
04Open source and portable.
The Last SaaS is open source, and the hosted version has no cost during the beta. Self-hosting is free: data exports with one command, and the server runs anywhere, from the start or after leaving the hosted version. Contributions and collaboration are appreciated and highly encouraged.
05Get started in three steps.
- Create an account
- Connect an assistant
- Apply the system to your particular use-case
Join the beta No cost during beta.
In Claude, connecting is Settings → Connectors → Add custom connector: paste the workspace's MCP address and approve the sign-in. Cursor and other MCP clients take the same address. The ChatGPT desktop app takes a custom connector the same way; on the web it currently requires a business workspace, and an app directory listing is planned. There's also a CLI.