Turn the systems you already run into services AI agents can use.
Enterprise applications, data platforms, content supply chains. We map what their data means and publish it as a governed MCP endpoint, the standard connection AI agents use. Claude, ChatGPT and your own agents then look up the real structure before they act.
Agents are already reaching into your systems. Most of them are guessing.
They guess what fields mean
The same customer is KUNNR in one system and Account__c in another. The agent fills the gap with what similar systems usually mean.
They invent values
Statuses, codes and categories your systems never accept. The call fails, or worse, it goes through with the wrong data.
They answer from the wrong version
A retired procedure, a draft policy, the wrong market's terms. Text search can't tell which copy is the one in force.
More prompting won't fix this. Agents need your real structure to look up before they act.
What we turn into services
For your own teams' agents, or offered to customers and partners as a service of your own.
ERP, CRM, HR, IT service, custom apps
- Check an account's real status before acting on it
- Raise a request with codes the system accepts
- Read only what that user is allowed to read
Warehouses, pipelines, catalogs, streams
- Answer from agreed definitions, not column guesses
- Catch a breaking change before it ships
- Write the data contract back to the pipeline
Documentation, regulated content, catalogs
- Retrieve the right step, version and audience
- Keep drafts and expired content out of answers
- Show where every answer came from
How it flows: one record, four stages
Follow a single customer record from an ERP extract to an agent's answer.
- 1 · Connect
We read the structure the system already publishes. No migration.
KUNNR: 00418
cust_stat: "A"
KTOKD: "Z001" - 2 · Map
AI proposes what each field means. Your people agree it and sign it off.
KUNNR ≡ Customer.id
"A" ≡ active
Z001 ≡ Wholesale - 3 · Transform
One model, written out in whatever format each system and agent needs.
JSON Schema · Avro
SQL · JSON-LD · OWL
tool definitions - 4 · Serve
The agent calls the endpoint and gets the record with its meaning attached.
get_customer("00418")
status: active
segment: Wholesale
Most AI search retrieves fragments of text. SCR retrieves the thing itself.
Structured Context Retrieval is our method for giving agents context. Instead of cutting documents into chunks and hoping the right ones come back, the endpoint returns whole units as your systems define them, with their type, their relationships, their version and where they came from.
Text chunks
typical AI searchWhich edition is current? Which rule applies to this account? The agent has to guess.
Structured Context Retrieval
your endpointNot today: two invoices are still open. The agent can say why and show its source.
We start from what you already publish
Your systems already describe themselves: a dbt manifest, a Salesforce org, a JSON Schema, a content model. We import that into one governed model, and every tool the agent calls is defined and checked against it.
When a system changes, the model changes, and the service follows. Nothing is hand-copied into a prompt.
Don't see yours below? If a system can export its structure, we can start from it.
See all 26 connectorsTen business days to your first service
- Days 1–2
Pick the system and the jobs
We agree on one system and the five to ten things an agent should be able to do with it, then import its structure.
- Days 3–7
Map and build
Your people confirm the mappings. We build the endpoint, read-only by default, with anything that writes on a separate connection.
- Days 8–9
Test with real agents
We run those jobs in Claude, ChatGPT and Cursor, and fix every place an agent guessed, stalled or picked the wrong tool.
- Day 10
Hand over
The code in your repository, the governed model, setup notes, and a test report your security team can read.
Fixed prices. No license to start.
Readiness test
Point us at one system, its schema, or an MCP endpoint you already run. We put real agent tasks to it and send a plain report of where agents fail, within 48 hours.
Fix
Your endpoint works, but agents struggle with it. We test it, ground it in a governed model, rewrite tools and descriptions, and retest. One week.
Pilot
One system, live as a governed agent service: mapped, built, tested against Claude, ChatGPT and Cursor, documented and handed over. Ten business days. Half to start, half at handover.
Each further system from $8,000. Upkeep as your systems change from $1,500 a month. Three pilot slots open for November.
Why us
We've spent years on one problem: getting systems, and the people who run them, to agree on what their data and content mean. CoreModels came out of that work, and so did Structured Context Retrieval. Our own company runs on agents that use MCP endpoints all day, so we know where they break.
Our methodology has been put to work at Adobe, Cisco, Eli Lilly, Mayo Clinic, Microsoft, PayPal, Sage Bionetworks and UnitedHealthcare.
Imported a dbt manifest with 8 datasets and 39 columns. Caught an allowed-values list that had quietly grown by one. Answered which models would break. Wrote the enforced contract back.
See the dbt exampleQuestions
Is this software we have to license?
No. It's a fixed-price engagement. You get a working service, its code and the governed model behind it. The model lives in CoreModels, and a year of a builder seat is included so your team can keep editing it.
Which agents will it work with?
Claude, ChatGPT, Cursor, and any other agent that connects over the Model Context Protocol, including ones your team builds. We test in the first three before handover.
Our system is custom, or old. Does that matter?
Usually not. If it can export its structure, as a schema, a database definition or a content model, we start there. A messy system is often exactly why agents struggle with it, and modelling it first is the part we're good at.
We care most about content, not data. Does this fit?
Yes. Structured content, such as DITA documentation, regulated content or a catalog with a content model, is where SCR does its best work. Agents get whole procedures and policies with version and audience attached, not fragments.
What about security?
Read-only by default. Anything that changes data sits on a separate connection that has to be granted on purpose. Every request runs as a signed-in user with that user's permissions.
Who owns what we build?
You do. The code is handed over in your repository, with notes your own engineers can maintain.
Find out in 20 minutes whether this fits.
Tell us which system and what agents should be able to do with it. We'll tell you plainly whether we can get you there in ten days.