Private MCP servers
Typed tools for search, calculation, file processing and controlled operations in company systems.
Service IT / ChatGPT plugins
ChatGPT plugins that work with your business systems
We turn ChatGPT or Codex into a practical interface for email, documents, CRM, ERP, databases, knowledge bases and internal APIs. The plugin does not just answer: it can find the right data, start an approved action and keep an auditable result.
What the product is
One plugin can combine instructions, tools and an interface
In the current OpenAI architecture, a plugin is an installable package for ChatGPT and Codex. It can include skills with repeatable instructions, an MCP server that exposes controlled tools and data, and an optional embedded interface. We select only the parts required by the business process.
What we design and build
From one focused tool to a complete operational workflow with maintenance.
Skills and operating rules
Stable instructions, domain knowledge, output formats, approval rules and concise operator workflows.
Interface inside the chat
Forms, status panels, previews and result cards when plain text is not enough for the task.
Business integrations
Email, Odoo, CRM/ERP, storage, databases, websites, messengers and external REST or SOAP APIs.
Authentication and access
OAuth 2.0, user- and workspace-bound access, minimal scopes and separation of customers and environments.
Testing and operation
Test scenarios, idempotency, logs, monitoring, versioning, publication preparation and ongoing support.
Connected sources
The chat becomes a working window into the existing process
The plugin can read, search and—where permitted—act across the systems your team already uses.
Operating logic
A clear route from request to verified result
We define the data boundary, tools and approvals before the model is allowed to act.
Trigger
A person asks in ChatGPT, or an approved event arrives from email, a form, CRM or a scheduler.
Understanding
The assistant identifies the task, required context, risk level and missing input.
MCP tool
A typed tool fetches or changes only the data allowed for that user and workspace.
Business rule
Validation, deduplication, routing and approval logic run on the controlled backend.
Answer or action
ChatGPT returns the result, asks for approval or performs the permitted operation.
Audit and follow-up
The system keeps identifiers, status and traceability so the conversation can continue safely.
Anonymized client case
A diagnostic report becomes a separate, actionable conversation
For one of our clients in automotive service, we implemented a workflow that receives reports from a professional diagnostic scanner, structures the findings and delivers a concise diagnostic route in ChatGPT.
The scanner sends a report
A monitored mailbox accepts only messages that match the diagnostic-report rules and passes the attachment to processing.
The report is decoded
The service extracts the vehicle identification number, mileage, control units, fault codes, states, freeze-frame and available live data.
The result is saved once
A private OAuth-protected MCP tool writes the normalized report idempotently, so a repeated mail event does not create a duplicate.
Cases do not mix
The context is keyed by the vehicle identifier and the current fault-code signature. The same combination stays in one case; changed faults start a separate branch.
ChatGPT gives an exact diagnostic path
The answer starts with the precise scanner menu path, then lists concrete checks, expected readings and stop conditions—without generic filler.
The technician continues in chat
They can ask for the next check, compare reports and request a repair procedure. A full vehicle procedure is given only for a verified variant and a trusted source.
Security and control
The model sees a tool—not unrestricted access
We place business rules on the server side and limit what every tool can read or change.
- OAuth and workspace-bound authorization
- Minimal scopes and role-based access
- Separation of customers, development and production
- Idempotency and duplicate protection
- Human approval for significant actions
- Audit logs, monitoring and rollback paths
- Data minimization and retention rules
- Separate AI Act/GDPR review for sensitive or high-risk use cases
Deliverables
A production-ready integration, not only a prompt
- Use-case map and tool specification
- MCP server and connectors
- Skills, prompts and output contracts
- OAuth and access-control design
- Optional embedded UI
- Automations and event routing
- Tests, monitoring and technical documentation
- Deployment, publication preparation and support
How we start
Process review
We select one repeatable task with clear data, result and owner.
Safe prototype
We test the dialogue, tool schema and failure cases without broad production access.
Integration
We connect the required systems, authentication, logs and approvals.
Launch and improve
We verify the real workflow, monitor quality and extend the plugin only where useful.
Questions before development
Is this a chatbot?
No. A plugin may use the familiar chat interface, but its value is controlled access to real tools, data and workflows.
Can it be private?
Yes. We can build a private company plugin with OAuth and restricted access, or prepare a suitable product for the OpenAI publication process.
Can it write to CRM or send an action?
Yes, if the action is explicitly designed, authorized and audited. Significant or irreversible actions remain behind human approval.
How long and how much does it cost?
That depends on the number of systems, authentication, data quality, UI and risk controls. After a short process review we provide a scoped estimate.
Next step
Show us one process that currently jumps between chat, email and business software
We will map the data route, identify which MCP tools are needed and propose a controlled first version.