Daslab Gateway

Your data does not have to move

Daslab Gateway is a single binary on a machine inside your network. It makes your ERP, your documents and your internal systems into things an agent can actually use, over one outbound connection.

You open no inbound port and you create no database user.

The connection runs the other way.

Your network
ERPFile sharesScanned documentsDatabasesInternal APIs
Gateway node One binary. Read-only. Holds the copy and answers against it.
Daslab — agents and scenes
The usual dead end

Why this is usually where it stops

The useful data is behind a firewall and the tools are not. That leaves three options, and everyone in the room already knows all of them.

The gateway takes the fourth.

01
Copy it to the cloudFast to say, slow to clear. The systems worth asking about are the ones whose data is hardest to get permission to move.
02
Open a path inwardA VPN, a jump host, an allowlist. Each is a project with a security review attached, and it ends with an inbound door somebody now owns.
03
Do nothingWhich is what usually happens, and why the question gets asked once a quarter in a meeting instead of once an hour by whoever needs it.
04
Leave it where it isThe node dials out and holds a read-only copy. A query from a scene never reaches your source system directly.
Purchase order 4500019283 in your scene
Fields — queryable today
VendorMeridian Steel
Net value128,400
Created2026-03-11
BuyerK. Alvarez
Attached — never a field
PDFQuote comparison sheet
XLSXSupplier pricing, 3 bids
PDFSigned amendment
JPGGoods receipt photo
The node stores each file next to the object it belongs to, so one question can span both.
The part SQL cannot reach

Your purchase orders have more than you can query

An ERP records what you bought and what you paid. The part that explains the decision is usually attached to it rather than typed into it: the quotes you compared, the specification you sent, the certificate that came back, the amendment somebody signed.

Those attachments are not reachable from SQL at all. In SAP the content sits in a binary column the query interface will not return, which is why every replication tool in this category moves rows and leaves the files behind.

Your agent reads the document itself. There is no extraction step to configure and no template to maintain, because a model reads a scanned comparison sheet the same way a person does.

"Show me every order over 100,000 where we took the higher of two quotes, and what the losing quote said."
TablesDocuments
Load and evidence

What it costs the system it reads

A ceiling you set, not a promise we make. The node holds itself to a fixed number of requests per second across all its workers and a fixed number of objects per run, so an unattended job cannot turn into a load event on a production system at four in the morning.

Reading an SAP system needs no database user and no database port. The node asks through the application layer, as a read-only account, and every statement it sends is a SELECT. There is no code path in it that writes.

Every read is recorded: the statement, the account that sent it, the row count, the time it took. That is the log your Basis team asks for first, and "we do not keep one" is the wrong answer.

The trail runs both ways. What the node asked your system, and what it answered back to a scene, are the same record. Every workflow becomes a scene — its data, its tools, the agent that runs it, and the history of everything it did — so a question asked six months ago still has its answer attached.

Trailboth directions
SELECTekko where ebeln > '45000192…'5,000412ms
SELECTekpo where ebeln > '45000192…'5,000508ms
GETattachments for 45000192834220ms
SELECTsrgbtbrel where brelguid > '9F2…'5,000301ms
ANSWERto scene "commodity margin" — 3 rows, 1 document88ms
4 requests / second 500 objects / run read-only account
$ curl -fsSL https://releases.daslab.dev/gateway.sh | sh
$ daslab-gateway connect
  paired with your workspace
$ daslab-gateway add sap --host erp.internal
  70 tables, keys and dates from the dictionary
  attachment service found
Install

What it runs on

One file. No runtime to install, no database to administer, no agent framework on the box.

The third command asks the system what it holds. Table keys and date columns come from the data dictionary, attachment services from the service catalogue. Nothing about your estate is typed in by hand, so nothing about it can be typed in wrong.

From there the node appears in your workspace as an asset. Any scene you give access to can query it, and every question and answer lands in that scene's history.