AI Access Layer · Custom Connectors
Your systems,
in plain language.
Under control.
We take your organisation’s own APIs and turn them into clear tools your people and AI agents can use. Read first. Writes wait for a person. Everything is logged.
Step 1 of 4: Request
What is the balance on loan LN-104472?
The request starts in plain language, with a named person.
Text version
Read: Amina W., Credit Officer, asks for loan LN-104472. The request maps to get_loan_balance, the role permits a read, and the sample answer returns KES 24,500.00 with national ID redacted. Audit entry 000412 records it.
Write: the agent requests a repayment-date change from 15 to 30 November 2026. The role check permits the request, a separate Credit Supervisor approves, and only then is the change shown as complete. Entry 000413 records before, after and who approved. These are invented examples.
Five steps. We do the building.
- 01
Understand your system
- You provide
- API access in a test environment, docs, a system owner.
- We deliver
- A map of the jobs people actually need done.
- 02
Design the tools
- You provide
- Sign-off on each tool and who may use it.
- We deliver
- One clear tool per job, in plain language.
- 03
Build and test
- You provide
- Sample questions from real staff.
- We deliver
- Tools tested against real model calls.
- 04
Permission and approve
- You provide
- Roles and who approves each write.
- We deliver
- Role rules and approval steps, live in the gateway.
- 05
Run and review
- You provide
- Feedback and change requests.
- We deliver
- Monitoring, audit reviews and updates as the API changes.
A clear tool gives AI less to guess.
An AI given a raw API guesses. A well-built tool doesn’t have to.
Raw API
- Hundreds of endpoints and one broad key
- Loose inputs
- Write access from day one
- Full records with sensitive fields
msharti tool
- One clear tool per job
- Typed inputs checked before the call
- Safe defaults, read-only first
- Only needed fields; the rest redacted
get_loan_balanceRead-only · scoped access- Does
- Returns the balance of one loan.
- Input
- loan_ref: a defined loan reference
- Mode
- Read-only
- Roles
- Credit Officer, within permitted scope
- Returns
- Balance and reference; national ID redacted
- Logged
- Who asked, the role decision and result
What a connector includes.
Tool definitions
Role permissions
Approval rules for writes
Audit events
Tests against real model calls
Handover and run
Read first, write with approval.
Read.
Start with a narrow, permitted view of the records needed for the task.
Write.
Tool by tool, only with human approval. Nothing changes before a person decides.
Logged.
Before and after, who asked and who approved. The result stays on the record.
Works with the AI your people already use.
AI assistants, AI agents and your own AI apps connect through one msharti gateway using MCP, the open standard for connecting AI to tools.
Before we build.
msharti builds it for your organisation. There is nothing for you to code. We scope the tools and permissions with your system owner and risk owner.
Your agreement and DPA define which fields are processed, what the audit log retains, retention periods, and how data is returned or deleted.
Only with human approval. Every connector starts read-only; write tools are added one by one under agreed role and approval rules.
A system with an API can be assessed. Feasibility and scope are confirmed during the assessment; each connector is built per client.
Book an assessment. We agree the system, the jobs people need done and who owns the risk before building.