UTEC · Final project
molu
An MCP front-end in Go that exposes the schema of a data system as tools for AI agents.
- Role
- Team development: hub, catalog and MCP transport
- Year
- 2026
- Stack
- GoMCPRESTSpec-driven development
The problem
molu is an MCP sidecar for xolu, a data system with business rules. It reads xolu's live schema and publishes it as Model Context Protocol tools, so an AI agent can operate with structure and with the guarantees xolu already gives. Alongside molu there is a hub where domain applications publish their own functions. It is the final degree project and is still in development.
What I built
- A concurrency-safe hub catalog and an API for publishers to register their functions.
- The hub's MCP transport and filtering of the offered tools through a configurable admission-and-score interface.
- Fixes in tool calls, in xolu's nested state transitions and in how list limits are handled.
- Scripts to bring xolu up on macOS, Windows and Linux, and memory and rules documentation for the team's agents.
Technical decisions
Spec-driven development
Each piece starts from a spec, a plan and a task list, and moves forward with TDD and race-detector tests as a gate. It is slower at first, but it let four people split the work without stepping on each other.
Typed errors that never collapse
Failures have six fixed codes, such as missing resource, invalid transition or system unavailable. The agent always gets a concrete reason instead of a generic error, which lets it correct its next step.
Pure logic kept apart from I/O
The xolu REST client, the MCP transport and the admission provider hide behind interfaces, so each part is tested with doubles without starting the whole system.
What I took away
- How the Model Context Protocol works inside: tools, discovery and calls.
- Working with AI agents as part of the team, with rules and memory written in the repository.
- Writing concurrency in Go and testing it with the race detector.