MCP servers
Let any AI assistant work with your systems.
We build MCP servers that connect Claude, ChatGPT, Copilot or your own agents to the systems you run. Each person gets answers from live data, with their own rights, and every change follows the approval rules you set.
See it on a system like yours
orders.search(status: "unshipped", placed: "yesterday")payments.list(due: "this_week")downtime.get(line: 3, period: "last_night")shipments.search(country: "DE", status: "late")licences.list(team: "mine", status: "unused")Why MCP
Build once. Every assistant can use it.
Your API is written for developers. An MCP server turns it into tools any AI assistant can find and use, with sign-in, limits and logging built in.
- Claude
- ChatGPT
- Gemini
- Microsoft Copilot
- Cursor
- Your own agent
An open standard, introduced by Anthropic and governed by the Linux Foundation since December 2025.
Without MCP
- Claude
- ChatGPT
- Your agent
- Shop
- ERP
- Ledger
Every agent needs its own integration with every system: 9 to build, secure and maintain.
With MCP
- Claude
- ChatGPT
- Your agent
MCP servers
- Shop
- ERP
- Ledger
One server per system, and any agent can use it: one place for sign-in, limits and the log.
Security and control
Agents get rights, not keys.
A demo server is easy. One that touches orders, money or a production line needs the controls your security review will ask for. Ours ship with them.
Acts on behalf of a named person
The agent has its own identity and works with that person’s rights, so every log line shows both: who asked and which agent acted.
Read-only by default
Every server starts read-only. Change tools are added one by one, each with its own limits.
A person approves changes
Beyond the low-risk actions you have allow-listed, every change waits for 1 approver, or 2 where you need four eyes.
Masked by policy
Your policy lists the fields to mask, and the server applies it, not the model. Card data is read only as tokens, never as full numbers.
Prompt-injection safeguards
No model is fully immune, so we limit the damage: the agent only reaches the tools and data you allowed, and changes beyond your allow-list wait for a person.
Every call on record
Who asked, which agent acted, who approved and what came back, in your log and SIEM, masked the same way as answers.
Runs in your infrastructure
The server sits in your environment, next to your systems. Your data does not pass through UAPP.
Kill switch
One step revokes an agent’s mandate or disconnects the server. Running tasks stop and no new calls go through.
How we work
One system first, then the rest.
We start with one system and the questions people ask most, prove the answers are right, and only then add change tools and more systems.
- Step 1Discovery
Questions and data map
We list the questions and tasks, the systems behind them and who may see what.
- Question list
- Data and access map
- Tool plan with owners
- Step 2Read-only
First server
One system, read-only tools, tested against your questions in your environment.
- Semantic data model
- Read tools
- Test results
- Step 32-4 months to a pilot
Changes and more systems
Change tools with confirmation, then more systems under the same rules and the same log.
- Change tools with approval
- One audit log
- Handover and documentation
Contact
Tell us which systems your agents should work with.
A 30-minute technical call with an engineer. Bring one system and the questions people ask about it; we tell you plainly what an MCP server for it would take.
- We look at one system end to end
- You get a data map and a tool plan
- Confidential from the first call, NDA if you need it
Prefer email? office@uapp.group