One definition per term, used consistently across the handbook, the console and the rest of the
site. Where the industry uses a word loosely, this page says what Msharti means by it.
Core terms
Term
Definition
Access plane
The half of the platform that connects AI assistants to business systems: the gateway, the connectors and the tools they expose. Called AI access platform on the marketing site.
Agent
A credential bound to a role bound to a tool allowlist. Not a separate object you create; it is the combination that makes an automated caller governed. See build an agent.
Allowlist
The list of tool names a role permits. Deny by default: anything not named is refused.
API key
A static bearer credential in the form msh_…, carrying a role. Identifies a machine rather than a person. Shown once at creation.
Connect Link
A one-time OAuth link Msharti generates for your tenant so you can connect a third-party system without registering an OAuth application of your own.
Connector
An isolated server that speaks to exactly one business system, such as M-Pesa Daraja or Salesforce. One container per connector.
Control plane
The half of the platform that decides what AI is allowed to do: identity, roles, injection detection, secret stripping and audit.
Delegated permissions
An assistant can only reach data the signed-in person could already reach. Connecting it never grants more than the human has.
Gateway
The single endpoint every assistant talks to, where authentication, authorisation, injection scanning, routing, redaction and audit all happen.
MCP
Model Context Protocol, the open standard that lets AI assistants discover and call tools. Msharti implements revision 2025-03-26 over Streamable HTTP.
Role
A named policy carrying a tool allowlist. Every credential has exactly one. Editing it changes what every holder can do, without reissuing anything.
Seat
A person in your tenant, carrying one role. Counted against your plan. API keys do not consume seats.
Tenant
Your account, and the isolation boundary around your connectors, credentials, people, roles and audit history.
Tool
One named operation an AI can invoke on a connector, such as mpesa_collection_summary. The unit of permission and the unit of audit.
Security terms
Term
Definition
Injection detection
Scanning requests for prompt-injection patterns. The threat is content already inside your own systems, such as an email written to redirect an assistant that reads it.
Read-only enforcement
SQL connectors accept SELECT only, rejected at the gateway rather than only at the database user. A database guarantee, not a platform-wide one.
Row-level security
Database isolation ensuring a query issued for one tenant cannot return another tenant's rows, even if layers above were bypassed.
Secret stripping
Redacting keys, passwords and sensitive identifiers from a response before the assistant ever sees it.
Write tool
A tool that changes something outside Msharti: sending mail, creating a record, initiating a payment. Grant deliberately.
Status labels
Label
Meaning
Live
Shipped and in production use today.
Beta
Working and available, still being hardened. Expect rough edges.
Roadmap
Not built. Named as roadmap everywhere it appears, never presented as available.
Enterprise
Available on the Enterprise tier.
Words we avoid
Not this
Because
"Msharti is read-only"
It is not. Databases are read-only. M-Pesa can move money and Outlook can send mail. Say which.
"Grant access to M-Pesa"
There is no connector-level grant. Permissions are per tool, precisely because one connector spans reporting and payments.
"Msharti stores your data"
It is a pass-through by default. Retention is an opt-in add-on, per connector.
"Msharti is an iPaaS"
There is no workflow engine or orchestration layer. It is governed access for AI, which is a narrower and different claim.