AI Service Desk Autopilot brings tickets, server diagnostics, and activity reports into one technician workspace. Built for the techbold challenge at START Hack Vienna 2026, it won the challenge and the ElevenLabs special award for best use of voice.
From ticket to activity report
The agent pulls a ticket from Phoenix ERP and connects to the customer’s Linux server over SSH. It gathers context, tries to reproduce the issue, and proposes ranked explanations for the root cause. The technician chooses a hypothesis, adds a comment, or supplies their own.
The investigation continues in a single tool-calling loop. If a hypothesis does not hold up, the agent revises it. Before a fix runs, the technician reviews a plan with commands, validation steps, and a rollback approach. The agent then checks the result and drafts an activity report for review before submission to the ERP.
Approval before execution
A deterministic safety layer checks commands at both the tool layer and the SSH runner. Allowlisted reads can run automatically when the technician enables that option. Other commands pause for approval; denylisted operations cannot be approved at all. The layer also handles output redaction.
Technicians can edit or reject a proposed command, pause for a decision, stop the run, or use a manual SSH terminal alongside the agent. The language model does not decide its own permissions.
Voice control without a bypass
The ElevenLabs co-pilot lets technicians ask for status, steer hypotheses, review pending commands, and stop a run by voice. Spoken actions pass through the same approval gates as the interface. The co-pilot reads a pending command back and asks for explicit confirmation before approval.
The backend creates a short-lived signed connection URL, so the ElevenLabs API key stays out of the browser. The workspace also works without voice configured.
Implementation
The frontend uses React, Vite, and an xterm terminal. The Python backend integrates Phoenix ERP, runs SSH sessions through AsyncSSH, and manages the agent loop and approval state. Docker Compose starts both services; a mock Phoenix ERP supports offline development and tests.