kvm-pilot

Smart hands for your AI agents. A write-capable, multi-plane (KVM + BMC + SSH) MCP server for controlling physical machines — gated, verified, audited.

kvm-pilot lets an agent drive a headless box through POST, firmware, the bootloader, and an OS install with no agent on the target: it works at the pixel level through an IP-KVM (PiKVM, the GL.iNet GLKVM fork GL-RM1 / GL-RM1PE, BliKVM), at the structured-state level through a BMC (Redfish on iDRAC/iLO/OpenBMC, IPMI on BMCs that predate Redfish), at the firmware level through Intel AMT/vPro (a BIOS/POST/GRUB screenshot + power + SOL on a business laptop an HDMI-capture KVM can't see boot on), and over SSH once an OS is up. A pluggable vision subsystem reads a KVM screenshot and tells you what boot phase the machine is in — bios_menu, grub_menu, installer_progress, login_prompt, crash_screen, and so on — and a safety layer gates every destructive operation behind operator opt-ins and per-call approvals.

Vision runs on Claude or any local OpenAI-compatible VLM (LM Studio, Ollama, vLLM, llama.cpp). Point it at a model on your own GPU and the screenshots never leave your network and cost nothing per frame.

How it works

kvm-pilot runs a see → decide → act loop, and the screen is its only sensor: it pulls a screenshot from the KVM, a vision model classifies the boot phase, and kvm-pilot acts back through the KVM's keyboard and power. Because it works at the pixel level, there is no agent on the target — the same loop drives POST, firmware, the bootloader, and an OS install.

A real, unedited run against a GLKVM on the home fleet — an honest preflight, a headless snapshot (watch it wake the on-demand encoder), a gated dry-run power-cycle, and the boot console the agent actually saw:

Quickstart

One install gives you the whole product — the kvm-pilot CLI, the kvm-pilot-mcp MCP server, and the bundled Claude skill — nothing to clone. The current release line is a pre-release: install it with --pre, which is explicit and keeps selecting the beta line once a stable release ships (0.1.0a1 is yanked and much older than this README — don't pin it).

pip install --pre kvm-pilot                    # CLI + skill + MCP server + WebSocket events
pip install --pre "kvm-pilot[totp]"            # + 2FA / TOTP support (pyotp)
kvm-pilot install-skill                        # optional: put the bundled skill where Claude Code loads it

Driving a KVM from an AI agent (MCP)

claude mcp add kvm-pilot -s user \
    -e KVM_PILOT_PROFILE=mykvm -e KVM_PILOT_MCP_READ_ONLY=1 -- \
    kvm-pilot-mcp

KVM_PILOT_MCP_READ_ONLY=1 is the recommended first rung of the trust ladder — the agent can see everything and touch nothing until your hardware is verified. The Getting started guide covers credentials, Claude Desktop JSON config, sample prompts, and climbing the ladder. The server is published to the official MCP registry as io.github.DustinTrap/kvm-pilot, so registry-aware hosts can discover and install it by name. Agents: the repo root carries an llms.txt doc map.

Scripting from Python

from kvm_pilot import KVMClient
from kvm_pilot.vision import ScreenAnalyzer, make_backend

kvm = KVMClient("192.168.8.1", "admin", "secret")

# Classify the current screen with Claude (model auto-resolved at runtime)
analyzer = ScreenAnalyzer(kvm, make_backend("anthropic"))
print(analyzer.classify().phase)

# Or run entirely on a local VLM — nothing leaves your network
local = make_backend("local", base_url="http://127.0.0.1:1234/v1", model="qwen2.5-vl-7b")
analyzer = ScreenAnalyzer(kvm, local)

# Block until the box reaches the GRUB menu, then pick the first entry
analyzer.wait_for_state("grub_menu", timeout=120)
kvm.press_key("Enter")

For the latest unreleased tree:

pip install "kvm-pilot[totp,ws] @ git+https://github.com/DustinTrap/kvm-pilot"

CLI

kvm-pilot info     --host 192.168.8.1 --user admin --ask-passwd   # prompt (no echo)
kvm-pilot capabilities --profile homelab                 # what this driver supports
kvm-pilot snapshot screen.jpg --profile homelab
kvm-pilot --timeout 60 power-cycle --profile homelab --dry-run   # log, don't send
kvm-pilot eject --profile homelab                        # detach virtual media
kvm-pilot events --profile homelab --count 5             # stream device events
kvm-pilot watch grub_menu --profile homelab \
    --backend local --vision-url http://127.0.0.1:1234/v1 --vision-model qwen2.5-vl-7b

The CLI prompts for confirmation before any destructive action (power, virtual media — including uploads — keyboard/mouse injection, GPIO). Use --yes to skip prompts in automation, or --dry-run to log intended actions without sending them — dry-run short-circuits before the prompt, so it never blocks waiting for input. --timeout (HTTP per-request timeout) is a global flag and goes before the subcommand; watch keeps its own --timeout for the vision wait deadline.

Profiles like homelab live in ~/.config/kvm-pilot/config.toml. See docs/cli.md for the full command table (every subcommand, the capability it needs, and its gating), and docs/configuration.md for the config-file format, every KVM_PILOT_* environment variable, and the precedence between flags, env, and profiles.

GLKVM setup note: on GL.iNet firmware the PiKVM REST API is disabled by default (every /api/* call 404s, surfaced as a clear ApiDisabledError), and a firmware upgrade can re-disable it. Enable it in /etc/kvmd/nginx-kvmd.conf and pin the driver with --driver glkvm / driver = "glkvm" — full steps in the troubleshooting guide.

The tool surface, by plane

The same capability protocols span three actuation planes, so one agent workflow can mix pixels, structured BMC state, and shell access — with every destructive effect gated per class:

A machine with no KVM and no BMC at all is itself a first-class target: driver = "ssh" plus ssh_host (the machine's own address, #248) stands kvm-pilot on the recovery ladder's in-band rung — every device capability honestly absent, and the healthcheck saying plainly that there is no out-of-band recovery path there.

The canonical per-tool reference — annotations, effect gates, approval lifecycle — is the MCP server README; the CLI covers the full surface in docs/cli.md.

Status & maturity

Status: release candidate. GA is gated on validation breadth, not code: three of the seven device drivers (redfish, pikvm, blikvm) have never run against real hardware, only emulators. See the Hardware-Compatibility list for what has actually been exercised — and please add to it. (The exact version lives in the CHANGELOG; install with pip install --pre kvm-pilot.) The core paths have graduated from mocked-only to live-verified: a fleet of GL-RM1PE units has exercised snapshot/healthcheck/logs/power_state/virtual_media/info across two firmware lines — on V1.9.1 the run ledger that ships in the wheel puts info/logs/healthcheck/power_state at ga, snapshot at rc, and virtual_media at beta — derived from real runs, never hand-edited — and a Dell iDRAC6 has exercised the IPMI driver live end-to-end (power, boot-device, sensors, event log, SOL serial console). The paths that can hurt are hardened: transports never re-fire a destructive request, MCP approvals are signed single-use receipts with an audit trail, and every destructive effect — power, HID, media, boot-config, appliance, SSH, external writes — has its own operator opt-in gate. Recent betas added remote boot-device control (Redfish, IPMI, and in-band efibootmgr), Wake-on-LAN, an IPMI driver for BMCs that predate Redfish, a serial (SOL) console, mouse auto-calibration, and headless native-resolution GLKVM snapshots; kvm-pilot test-report turns contributing evidence into one command, and the firmware registry feeds itself (firmware-check auto-files registry updates). Now we need your hardware. PiKVM, BliKVM, other GLKVM models, and Redfish BMCs (iDRAC/iLO/OpenBMC) are the combos the matrix needs most — success or failure, a hardware report takes two minutes and the hourly ingest does the rest. Anything the Hardware-Compatibility list doesn't show as exercised is still unverified: expect some API movement before 1.0, note the remote firmware-flash no-op on GL-RM1PE (#94/#95), and don't point destructive ops at a machine you can't afford to have power-cycled unexpectedly. See Compatibility.

Boot-phase detection

The vision classifier maps each screenshot to a phase — bios_menu, grub_menu, installer_progress, login_prompt, crash_screen, and so on. wait_for_state() polls the screen and blocks until the phase you asked for appears (or a timeout fires), so an unattended install becomes a few waits with actions wired between them:

Sensing model

Vision is the most expensive way to read a screen — a model call per frame — and most of what it infers (power state, boot phase, liveness, a crash) is also available as a field, an event, or a line of text. The direction of kvm-pilot is to treat classification as a hierarchy: answer from the cheapest signal the device exposes, and fall through to OCR and finally a vision model only when nothing cheaper can.

The PiKVM/GLKVM client already exposes the cheap end — ATX and HID LEDs, video-signal and resolution, on-device OCR (?ocr=true), logs, Prometheus metrics, and a WebSocket event stream. The capability protocols add Logs, BootProgress, Sensors, SerialConsole, Watchdog, and BootConfig as the seam for BMC drivers (Redfish/IPMI), where the boot phase is a structured enum (BootProgress.LastState) and the console is a serial text stream rather than pixels. Different device classes are nearly complementary: capture devices are strong on pixels, BMCs on structured state and serial text.

Safety model

Power-offs, hard resets, virtual-media connect/disconnect and image uploads, keyboard/mouse injection (type_text, press_key, shortcuts, clicks), GPIO, boot-config changes, and Redfish/IPMI resets are classified as destructive and pass through a safety layer:

• dry-run short-circuits first: it logs the intended call and skips it entirely — the confirm callback is never invoked, so dry runs never prompt or block.
• confirmation — a callback that can veto any destructive call that would really be sent. The library default allows everything (so plain scripts work); the CLI installs an interactive y/N prompt unless you pass --yes.

The destructive set is defined explicitly in kvm_pilot.safety.DESTRUCTIVE_OPS so it is auditable rather than guessed. A vision classification can never trigger a destructive action on its own — you wire that yourself, and the safety layer still applies. On the MCP side each destructive effect class additionally needs an operator opt-in env gate plus a per-call approval backed by a signed single-use receipt — the trust ladder (READ_ONLY → DRY_RUN → per-effect ALLOW_*) is drawn in the MCP server README.

This software controls real hardware and can power-cycle or interrupt a running machine. Read SECURITY.md before exposing a KVM to the internet.

No hard-coded model version

There is no model version string anywhere in the code. The Anthropic backend resolves the newest vision-capable model at runtime via the Models API and caches it; set KVM_PILOT_VISION_MODEL or pass model= to pin one. The local backend uses whatever model you loaded on your server. Bring your own backend, endpoint, and model.

How this differs from other clients

pikvm-lib is a fine general-purpose PiKVM client. kvm-pilot is aimed at a different job:

• Vision-based boot-phase detection — classify BIOS/GRUB/installer/crash states from screenshots, with blocking wait_for_state loops. This is the core feature and pikvm-lib has no equivalent.
• Pluggable local or cloud VLM — run inference on your own GPU at zero per-frame cost, or on Claude.
• A safety layer around destructive operations (dry-run + confirmation).
• GLKVM-fork awareness — documents the API-enable prerequisite and GL hardware quirks that bite GL-RM1PE users.
• Stdlib-only client core — the driver/vision code imports only the standard library (the bundled MCP server pulls the mcp SDK; feature extras are opt-in).

If you just want to script power and HID against a stock PiKVM and don't need the vision layer, pikvm-lib may be the simpler choice.

On the BMC side, sushy, DMTF's python-redfish-library, and pyghmi (IPMI) are mature, far more complete BMC management SDKs — if you need account/firmware/network configuration, EventService subscriptions, or hardware-proven maturity, use them. kvm-pilot trades that completeness for one uniform capability surface across device classes (IP-KVMs and BMCs behind the same protocols), the same safety layer gating every destructive call, and the vision loop on devices that have pixels.

Compatibility

The GL-RM1PE (read/snapshot paths), a Dell iDRAC6 over IPMI, and a Dell Latitude 5411 over Intel AMT are the combos run live so far — everything else is "expected to work" pending validation. The Hardware-Compatibility list is the authoritative, per-capability record. ATX power control needs the ATX adapter wired to the target's front-panel header: on the GL Comet family (GL-RM1 / GL-RM1PE) that is GL.iNet's separately sold ATX board (GL-ATXPC), while PiKVM v3/v4 kits include the ATX adapter in the box and BliKVM bundles vary by model — check yours. Without ATX wiring, ATX calls return errors from the device. Reports of success or failure on any hardware are exactly what this beta needs — please open a hardware report.

Architecture

kvm-pilot is built on a modular, driver-plugin architecture so support can expand to many KVM/BMC devices (PiKVM family, Redfish BMCs, IPMI BMCs, JetKVM, …). Each device implements only the capability protocols its hardware supports; the CLI, safety layer, and vision subsystem stay device-agnostic. A make_driver(kind) registry (mirroring make_backend) builds drivers by name, and a hardware-free FakeDriver lets you exercise the whole loop — capabilities, safety gating, the analyzer — with no device (kvm-pilot capabilities --driver fake). See docs/architecture.md for the design and diagram.

A RedfishDriver (make_driver("redfish")) speaks the DMTF Redfish API to server BMCs — Dell iDRAC, HPE iLO, Supermicro, Lenovo XCC, OpenBMC — in one stdlib-only client. It shows why capabilities are segmented: a BMC's set is complementary to a PiKVM's (strong on structured state — power, boot phase, sensors, logs, virtual media — with no keyboard/mouse/screenshot), and the driver stays portable by following Redfish hypermedia rather than hard-coding vendor ids:

from kvm_pilot.drivers import make_driver

bmc = make_driver("redfish", host="idrac.lan", user="root", passwd="…")
bmc.get_boot_progress()        # 'os_running'  — structured, no screenshot
bmc.read_sensors()["temperatures"]
bmc.power_off(wait=True)       # mapped to the target's actual ResetType, gated

An IpmiDriver (make_driver("ipmi")) covers BMCs that predate Redfish (e.g. Dell iDRAC6) over the system ipmitool: power, boot-device control, sensors, the SEL event log, and an SOL serial console (kvm-pilot console). Both are on the CLI too — kvm-pilot info --driver redfish --host idrac.lan …. Capability-specific subcommands a BMC can't serve (type, snapshot, events) fail cleanly rather than crashing. Add --redfish-auth basic for an endpoint without a SessionService (emulators, or a BMC with session auth disabled).

An AmtDriver (make_driver("amt")) manages Intel AMT/vPro laptops and desktops out-of-band over AMT's three native protocols — WS-Man (power / info / single-use boot), SOL (kvm-pilot console), and KVM redirection — all pure-stdlib. It is the first non-PiKVM driver with Video + HID: snapshot returns a real firmware-level BIOS/POST/GRUB screenshot on a machine whose HDMI a capture-KVM can't see boot on a laptop. SOL and KVM listeners are opened remotely over WS-Man (kvm-pilot amt enable-sol / enable-kvm), no MEBx trip.

Help shape it — this is where you come in

kvm-pilot is built in the open, and the thing standing between it and 1.0 is not code — it's breadth of real hardware. Three of the seven drivers (redfish, pikvm, blikvm) have never run against a physical device, only emulators. Every rating you see is derived from a run ledger that ships inside the wheel, never hand-set, so the matrix only grows when someone runs it on real metal. That someone could be you.

The single most valuable thing you can send: a hardware report. kvm-pilot test-report --profile <name> turns it into one command, it is read-only by default, and an hourly job folds the result into the published compatibility matrix. Failures are as welcome as successes — a driver that returns a confident wrong answer on your BMC is a more useful report than one that works, and it will be treated as a first-class finding, not swept up.

Also genuinely wanted:

• ⭐ Star the repo if the idea is useful to you. It is the cheapest signal that this problem is worth solving, and it is what brings in the operators whose hardware the matrix still needs.
• 🐛 Open an issue for anything: a bug, a device that misbehaves, a confusing error, a doc that lied to you. This repo is issue-per-finding — an issue is the unit of record, and "the error message didn't tell me what to do" is a legitimate report.
• 💬 Tell us what it got wrong. Review the safety model, the approval flow, the tool surface. Push back on the defaults. If an agent did something with your machine that surprised you, that is exactly the feedback that makes this safe for everyone else.
• 🔌 Ask for your device. Missing driver, unsupported BMC, a KVM we've never heard of? Open a driver request — the plugin architecture exists precisely so that adding one doesn't mean forking the project, and docs/plugin-development.md is the contract.
• 🛠️ Send a PR. docs/CONTRIBUTING.md has the full pre-PR checklist; good first issues are labelled.

Nothing here phones home. The project has no telemetry that reports itself — every row in the compatibility matrix is there because a human chose to send it. That's the deal, and it's why your report actually matters.

Documentation

Full user and developer docs live in docs/ (architecture, design decisions, the Redfish reference, the troubleshooting & FAQ, contributing, and the security policy). The project wiki is an auto-generated, nicely formatted mirror of that folder, and the repo root carries an llms.txt doc map for AI agents.

License

Apache License 2.0 — see LICENSE and NOTICE. kvm-pilot is independent and not affiliated with or endorsed by the PiKVM project, GL.iNet, or Anthropic; those names are used only for compatibility description.