Agentic Fleet Router
Natural-language instructions become routing constraints in a moving AGV fleet — using Model Context Protocol as the bridge between operator intent and Dijkstra-based fleet routing.
What it does
An operator types a plain-English instruction — “block AGVs from aisle 3 for the next ten minutes” — and the system translates that intent into a constraint on the live routing algorithm. The constraint takes effect immediately; the next AGV asked to route through that zone gets a path that goes around it.
The interesting part is how. There’s no rule engine, no DSL, no chatbot. There’s an LLM that has been handed a small, well-typed set of tools via the Model Context Protocol (MCP). The LLM reads the operator’s instruction, picks the tool, calls it with the right arguments, and the tool mutates the routing graph’s edge weights. Dijkstra never changes — only the inputs do.
Why it matters
Manufacturing-floor operators know the floor. They don’t know graph theory, and they shouldn’t have to. This pattern — natural-language instruction in, structured operations action out — generalises across the floor: rescheduling production lines when a machine drops, rerouting around a maintenance window, prioritising one work order over another. Same paradigm, different toolset.
Architecture
- gRPC is the only wire to the fleet management system. No MQTT, no REST.
- Three layers, strictly separated: gRPC client → domain services (where the routing-graph logic lives) → MCP server (exposes tools to the LLM).
- Dijkstra is a black box. We modify edge costs and node constraints; we never touch the algorithm.
Try it now
Open the live demo and send the agent an instruction.