Leptex AI Gateway
Connecting merchant-authorized AI assistants to Workspace, with clear data access, action permissions, and integration checks.
Leptex AI Gateway is an AI integration capability within Workspace. Here, a gateway is a controlled channel through which external assistants access a merchant's business system. It lets an external assistant chosen and authorized by a merchant read customer conversations, review orders, and help perform permitted business actions. The engineering challenge is to connect an assistant's language capabilities to real business operations while keeping information access, authority, and results clear.
This article is for developers, integration teams, and business owners evaluating that connection. The current implementation uses the Model Context Protocol (MCP); client compatibility and availability still require separate confirmation. We describe the engineering capability and its limits, rather than treating an implemented interface as proof that every assistant is compatible.
Start with one pending task
Lucky Farm is the first customer onboarded onto Leptex. Its hawthorn business gives this connection concrete questions to address: customers ask about pickup or delivery, while the merchant needs to distinguish inquiries from orders and identify matters still awaiting a decision. We use that type of farm work to explain engineering requirements; this does not imply that Lucky Farm has enabled external agents or automated order actions.
Farms are also our initial focus for engineering validation. An operator may be harvesting, selling, and arranging collection at the same time; receiving external assistants should not require maintaining another complex system. Forward Deployed Engineers need to establish which questions existing information can answer and which requests need the owner, then use operating feedback to test whether an integration reduces the total burden. Availability and effects of each integration need separate confirmation.
For example, an owner may want to organize which customers are waiting for a reply and which orders need checking. An assistant needs to answer from the original records it may access. A customer asking to reschedule in a conversation must not become a summary saying that the farm has agreed to the change.
Handling that distinction well in a farm workflow provides a basis for assessing other settings. Fulfillment terms may differ between businesses, but the starting point remains the same: check pending work against the original records, then determine who needs to decide.
Connect the assistant to the business system
The connection has three roles:
- The AI assistant interprets the owner's request, reads permitted business information, and proposes or performs relevant actions.
- Workspace maintains customer conversations and order records, applying existing business permissions and state rules.
- AI Gateway provides a controlled access channel for the external assistant, keeping calls within the merchant's authorization.
The Model Context Protocol architecture describes a public protocol through which AI applications exchange context with external services and call tools. MCP provides a connection mechanism; prices, supply terms, order status, and authority still belong to the business system. Adopting a protocol does not complete a business integration on its own.
Leptex CaaP (Chat as a Page) is the merchant's customer-facing product. AI Gateway is Workspace's integration capability for authorized assistants. They support different parts of the same business. Here, gateway means access to business systems, rather than a general-purpose model API routing service.
Separate reading from acting
A connection should begin with the access required for its task. The current implementation defaults to read-only permissions; additional capabilities require separate authorization.
Swipe to view the full comparison →| Business need | Access and effect |
|---|---|
| Understand pending work | Read conversations and related order records by default, including items awaiting attention |
| Consult business information | Read project knowledge with additional permission |
| Help handle conversations | Reply to customers, take over a conversation, or hand it back with conversation-action permission |
| Handle order cards | Confirm or decline pending order cards with order-action permission, subject to existing business-state rules |
Reading information and changing business state carry different responsibilities. Replies appear in the conversation as the merchant; order actions change records in Workspace. Before enabling either, the owner should identify the work they intend to delegate.
Permission to act does not mean the assistant will ask for approval every time. The current connection does not add a separate approval step before each reply or order action. An instruction such as "ask me first" is not a substitute for that control. Owners who want to decide each action personally can keep the connection read-only and handle replies and orders in Workspace.
Begin integration with read-only checks
For accounts where connection settings are available, we recommend the following validation sequence:
- Review the connection settings in Workspace, select the minimum read-only access, and create a connection credential.
- Configure the chosen assistant through its supported integration process. Keep the credential in its supported credential-management facility, outside public documents, screenshots, and shared prompts.
- Query pending work and compare it with the original records in Workspace. Confirm that the connection accesses the intended business information and that the summary adds no commitments that were never made.
- Validate necessary additional permissions in a controlled test scenario. Check that action results match Workspace. If the business state does not allow an action, the failure should be explicit; an attempted action must not be reported as completed.
Reaching the interface is only the beginning. An integration's usefulness also depends on complete information, appropriate permissions, clear failures, and results the owner can understand.
Make security boundaries inspectable
The current implementation provides authorization scopes, credential expiry, revocation, and request-usage controls. Business actions remain subject to existing permissions and state rules. These controls constrain what the connection can do; they do not prevent every misunderstanding or guarantee protection against every attack.
A customer's words are business information to interpret, not instructions granting the assistant additional authority. The interface should enforce the relevant access and business constraints, and integration testing should examine how the assistant responds to such content. The MCP security best practices provide a public reference, not a substitute for validating a particular system.
Read-only connections also share data. Owners need to understand which conversations or customer details an external assistant will receive and how that service uses the information. Revoking a connection prevents further access with that credential; it does not automatically delete data already received by an external service.
The engineering goal is for an ordinary business owner to understand three things: what the assistant can see, what it can change, and how to stop further access. This translates the argument for knowledge boundaries and human authority in Beyond Personal Agents, Toward Personal Intelligence into requirements that can be checked for this integration.
Distinguish implementation, compatibility, and availability
The connection interface is implemented, and Workspace includes settings and guidance for Meta Muse. Compatibility with particular clients, accounts, and real business operations still needs individual validation. This does not imply an official Meta partnership or certification.
MCP support does not mean every assistant will work without further integration. Connection methods, authorization support, tool behavior, and error handling still need checking. Scheduled checks and notifications also depend on the external assistant's capabilities; the current connection itself does not promise continuous monitoring or instant alerts.
This interface is for merchant-authorized assistants. An agent asking for a quote on a buyer's behalf does not gain access to the merchant's internal information or order controls by visiting the public storefront. Broader reception and routing of buyer-agent requests, and support for additional assistants, require separate design and validation.
Validate business results
Integration acceptance should cover at least these outcomes:
- Reading: the assistant finds the correct conversations and orders, and its summaries can be checked against original records.
- Authorization: information and actions outside the granted scope remain unavailable; expired or revoked credentials cannot continue accessing the system.
- Actions: permitted changes appear correctly in Workspace, while invalid state transitions are rejected.
- Failures: the assistant distinguishes success, failure, and an outcome that is still unknown. It does not report uncertainty as completion; records should be checked before deciding whether to retry.
- Usability: the owner can understand permissions, inspect results, and end a connection they no longer need.
These are acceptance requirements, not a claim that this article documents completed testing across all clients. Evaluating value also requires observing handling time, correction effort, and missed work. More requests or fewer notifications alone cannot establish an improvement in business operations.
Ground collaboration in verifiable business records
An AI assistant participating in a business needs reliable information and clear authority. Leptex AI Gateway starts with Workspace integration, making conversations, orders, and permitted actions available to the owner's chosen assistant. Engineering progress should make tasks easier to complete, information easier to verify, and responsibility for resolving problems clearer.
CaaP helps customers understand a business, Workspace supports its operations, and AI Gateway lets an authorized assistant participate. Our integration path is to validate reading, authorization, execution, and failure handling for one task before expanding the scope of collaboration.
Learn more about Workspace, or contact us to discuss integration scope and your use case.