Work / AI engineering demo
Helix
Tool calls you can see.
A small read-only tool loop: intent, a router, Zod-validated arguments, then a visible UI effect. The loop plans, selects one tool, observes, and stops.
- Problem
- Agent demos often imply a large tool set without showing the call boundary.
- Role
- Independent engineer. Public demo on this site, not an employer case study.
- Implemented scope
- Two read-only tools, typed arguments, a visible router, and an explicit stop. A public write path is an abuse report, not a product tool.
Architecture
- 01
Intent
The visitor states a task. The demo treats that text as the only input to the loop.
- Responsibility
- Capture a bounded task
- Concern
- No hidden goal
- 02
Router
The router chooses one read-only tool. It does not open a general tool zoo.
- Responsibility
- Select one tool
- Concern
- Explicit stop
- 03
Zod-validated tools
Arguments are checked before a tool runs. A public write tool is an abuse report.
- Responsibility
- Reject bad arguments
- Concern
- Typed boundary
- 04
UI effect
The result is shown in the interface so the call is visible rather than implied.
- Responsibility
- Show the observation
- Concern
- Inspectable agency
What holds
- Decision
- Fewer tools, so each call stays inspectable. This is not a hosted MCP server.
- Failure handling
- Invalid arguments fail validation. The loop stops instead of continuing with an untyped call.
- Evidence
- The lab on this site runs a Helix-shaped plan, select, execute, and verify sequence and labels it as an interactive demo.
- Status
- AI engineering demo
- Limits
- Hosted MCP is learning, not a production server claimed here. No usage metrics are published.