AI Service Desk Autopilot

An AI-assisted technician workspace for the techbold START Hack track. It pulls tickets from the Phoenix ERP, connects to a customer Linux VM over SSH, and diagnoses + fixes the incident under the technician's control using a Cursor-Debug-Mode-style loop, then writes a clean activity back to the ERP.

Tickets list

Ticket detail

Command approval

Requires Docker and Docker Compose.

cp .env.example .env
cp /path/to/your-keys.pem keys/case{n}_key.pem # one key per ticket case

Minimum for Builder Base (real Phoenix + SSH):

PHOENIX_API_BASE_URL=https://your-phoenix-url-from-builder-base
PHOENIX_API_TOKEN=your-team-token
SSH_PRIVATE_KEY_PATH=/keys/case1_key.pem 
OPENROUTER_API_KEY=your-openrouter-key 

Ticket 700N uses keys/caseN_key.pem (N = ticket id − 7000). Put keys in ./keys/ (Docker mounts it read-only at /keys). SSH_PRIVATE_KEY_PATH can point at any key in that folder — the backend picks the matching sibling per ticket.

docker compose up --build

• Frontend: http://localhost:5173
• Backend: http://localhost:8000/health

One continuous autonomous agent (tool-calling loop), not a fixed pipeline!:

New Ticket:
-> Start: SSH connection
-> Automatically collect context with a standartised set of commands
-> AI tries to reproduce the issue
-> AI creates different ranked hypotheses
-> Technician picks what hypothesis to verify
-> AI verifies the Hypothesis by investigating
-> If root cause is not correct: create new hypthesis
-> AI proposes fix
-> AI collects proov to validate the fix
-> AI drafts an activity report
-> Technician reworks report and updates the ERP-System

The agent drives the investigation, but the technician steers at every gate:

• Hypothesis pick — after reproducing, the agent presents 2+ ranked root-cause hypotheses; the technician selects one, writes their own, or leaves comments that are feedback as steering guidance
• Command gate — state-changing commands pause for approval, the technician can approve, edit, or reject (with an optional reason)
• Fix review — ProposeFix is presented as one reviewable plan (explanation, commands, validation, rollback), the technician can influence the agent also here
• Decision pause — when blocked or uncertain, the agent calls RequestDecision and waits for the technician input
• Activity sign-off — the agent drafts the ERP activity; the technician reworks it before submit
• STOP / manual shell — STOP aborts the run immediately; the technician can also drop into a plain SSH terminal alongside the agent

• all requests are sorted into the Following categories
  • Safe Reads - a whitelist of commands that are recognized as "safe" and autoapproved when "autoapproving safe reads" is on
  • Deny — destructive operations (recursive rm, DB drops, firewall off, secret reads) are blacklisted and cannot be run such that even the technician cannot approve it
  • Confirm - All other commandshave to be approved by the technician
  • Redact - if a command redacting secrets is confirmed nevertheless the redaction feature delets all exposed features

A hands-free co-pilot built on an ElevenLabs Conversational AI agent. The technician can steer the entire run by voice while keeping their eyes on the terminal — every voice action goes through the same human-in-the-loop gates and safety rules as the UI; voice is just another input, never a bypass.

The agent can:

• Read run state — "status" summarises the phase, pending approval/decision, ranked hypotheses, and whether an activity draft is ready
• Steer hypotheses — select, comment on, or dictate a brand-new hypothesis
• Approve / reject — reads the pending command back and asks for explicit verbal confirmation before approving (safety gate)
• Answer decision prompts, toggle auto-approve for safe reads, send a one-shot command (through the normal mutation gate), submit the activity draft, or stop the run
• Speak up proactively — the frontend pushes contextual updates so the agent narrates new approvals, decisions, hypotheses, draft-ready, and phase changes without being asked

Security: the backend mints a short-lived signed wss:// URL from ELEVENLABS_API_KEY + ELEVENLABS_AGENT_ID — the API key never reaches the browser. When those vars are unset, voice degrades gracefully and the rest of the workspace is unaffected.

Full setup (dashboard agent config, client-tool definitions, system prompt) is in docs/voice-agent.md.

frontend/ React + Vite — TicketList, Workspace, xterm terminal
backend/app/
 erp/client.py Phoenix REST client
 ssh/runner.py asyncssh runner
 safety/rules.py denylist + mutation gating + redaction
 agent/loop.py autonomous tool-calling orchestration
 agent/session.py conversation + context compaction
 runs/manager.py run registry + human-in-the-loop gates
 mock/phoenix.py offline ERP for dev/tests

The LLM is never the security boundary — safety/rules.py enforces hard denies at the tool layer and again at the SSH runner.

MIT