Channels and roadmap
Bootscry aims to give an agent the same practical feedback a person needs: eyes, hands, and, where hardware allows, ears. The design gives each channel a driver on a shared interface. The statuses below describe working code and physical evidence as of September 2026.
Observation and action channels
| Channel | What it provides | Current status | Next milestone |
|---|---|---|---|
| Video KVM: GL.iNet Comet | Screen frames, keyboard, mouse | Working reference on one physical test bench; partial certification | Complete recovery and multi-target checks, then publish limits |
| Other video KVMs | Screen and input on vendor hardware | No driver or physical verification yet for PiKVM, JetKVM, NanoKVM, TinyPilot, or Redfish consoles | Obtain evaluation units, build and test drivers, publish per-device evidence |
| Serial console | Exact console text and input, without OCR | Planned driver; the test path starts with a pseudo-terminal pair | Implement read, write, and wait-for-text, then test on an owned disposable guest |
| Camera | Views of a screen, LEDs, or rack; possible microphone and speaker | One camera is activated and isolated in the lab. No agent driver, audio loopback, or PTZ control is verified | Complete local media setup and loopback, then prototype snap, clip, listen, say, and supported PTZ actions |
| HID-only input | Keyboard and mouse without a video feed | The Comet input path works; a standalone HID-only driver is planned | Pair input with a live serial, camera, or video observation channel; never run blind |
| Server management | Potential text console, virtual input, and power controls | iDRAC, Intel AMT, and Redfish are research topics only; no management driver has been tested | Design a capability map; obtain explicit operator authorization before management-interface testing |
Shared work above the drivers
Reading and phase awareness
Local OCR works on measured Comet frames. A local vision fallback and a classifier for BIOS, GRUB, installer, and login are planned. The vision result so far is an offline four-frame comparison.
Safety across channels
Read, input, and power tiers; dry-run; audit logs; stale-observation checks; and treating screen or console text as untrusted input are planned. See the safety design and current limits.
Agent tools
An MCP server with tier-gated tools and a fresh observation after input is planned. A build-versus-contribute decision is due in October 2026.