How the system fits together

lmkgpt is a single pipeline. You describe a job once. The app analyzes the request, runs a coordinated set of model calls, and stores enough state for you to audit what happened. The diagram below is the through-line most visitors need before they read individual features.

What else sits in the path

The high-level flow is the spine. These pieces sit alongside it. Each one exists to keep sessions honest, safe, and inspectable.

• Authentication and guest sessions. Signed-in users and guests both run missions; identity gates what is saved and what trial limits apply.
• Guest trial. Usage rules limit free runs so the product stays sustainable without blocking the core try-before-you-buy path.
• Moderation and prompt safety. Requests pass automated checks before plans are built; blocked prompts never reach the model layer.
• Orchestrator and sub-agents. A coordinator model issues a plan; worker models run in parallel, then a synthesis step merges results.
• Specialist routing. Domain capability packs can dispatch work to the right tool or runner instead of a single generic path every time.
• Mission knowledge. Structured knowledge units and optional vector search connect prior missions to new ones when the product needs recall.
• Company directory. A published directory lists organizations and suggested missions; it is a guide, not a live substitute for a fresh plan.
• Admin. Operators can manage internal metrics, profiles, and directory data. End users do not need this surface.

Further reading

These are independent references, not marketing copy. They are the sort of sources a serious buyer or architect expects when evaluating agent systems, streaming UIs, and database-backed knowledge.

• Building Effective AI Agents (Anthropic)
• Using server-sent events (MDN)
• Server-sent events (WHATWG HTML)
• Retrieval-Augmented Generation (Lewis et al., 2020)
• pgvector