AI that works inside your operations

AI profiles for the multi-tenant MSP

Each AI profile is a complete configuration: which model provider to use, whether chat and embeddings are enabled, which agents belong to the tenant, and how the profile authenticates. Profiles are identified by an AI key, with optional OAuth2 client-credentials (the AI key as client_id and a secret key as client_secret) so other systems can call in safely.

Access is first-class platform security. Keys can be revealed only when needed, configuration changes are permission-gated, and you can inspect token usage across user chat, ticket agents, monitoring agents and embeddings.

Bring the model that fits the customer

Cloud providers include OpenAI, Google Gemini, DeepSeek, Qwen and Aliyun, with encrypted API keys and configurable base URLs and request timeouts.

For customers who need data to stay on-premises, Ollama is a first-class local provider: hostname, port, HTTPS, parameter size, context length, and administrator overrides for context windows. Chat and embedding models are selected independently, so a compact embedding model can sit beside a larger reasoning model.

Agents that already know your platform

Service desk agents include a Ticket Template Assistant for structured ticket work, and an NMS Ticket Assistant that reasons over monitoring-driven tickets.

Monitoring agents analyse syslog streams, Windows event logs, and harvested device configurations — with batch size controls so you decide how much log context each run consumes.

A User Chat Assistant sits with operators in the product, using the same skill, tool and MCP configuration as the specialised agents. Multiple agents can be assigned to the same event and run in a defined order.

Embeddings for tickets and knowledge

Enable embeddings on a profile to index tickets and knowledge articles for retrieval. Vector storage uses Microsoft SQL Server 2025, or PostgreSQL with pgvector — aligned with the database estate you already run.

Embedding usage is tracked separately from chat. You can rebuild by deleting ticket or knowledge-article embeddings when the model or corpus changes.

Skills, function tools and C# event scripts

Skills are reusable bundles of tools. Child skills can be nested, and the tools an agent actually sees are filtered to those that apply to that agent type.

Each agent has a system prompt (with SmartFormat templating), default or custom function-tool sets, and C# scripts for OnAIExecuting and OnAIExecuted — plus user and assistant message hooks for chat. Scripts can cancel further processing, pass properties between agents via an AI Event object, and request structured integer, string, boolean or number outputs from the model.

The same scripting fabric used elsewhere in Gemmb is available here: HTTP and email clients, device live queries, SOAP services, timers and item bags, so an agent can act on infrastructure rather than only describe it.

MCP server and MCP clients

Gemmb can publish an MCP server so external AI clients use your operational tools. Authenticate with OAuth or an AI key in an HTTP header at a stable URL, or let end users bring their own API key on a tagged URL that selects the correct AI profile.

The platform is also an MCP client: register remote MCP servers, authenticate with API keys or OAuth client credentials, discover tools and resources, and assign those clients to individual agents.

Operable by design

Every profile and agent records token usage. Chat history and tool-call payloads are available for review. Profile logs include messages, timestamps, log data and stack traces, with optional debug logging and error-notification email.

Configuration can be validated before save, so a broken prompt or tool set does not silently reach production.