The platform, feature by feature.

Windows 7 through 11 and Server on x86 and x64; current Linux servers; old-glibc NAS appliances; fully static embedded builds for QNAP, Synology and OpenWrt-class targets down to ARMv5; macOS on Apple silicon and Intel. One agent line, one protocol, one console — the build profile is a decision about the host, not about the tooling.

Every customer build is structurally unique. That is a property of the toolchain, not a setting you remember to enable.

An alternate Windows implant with the full command set resident in the binary. Same server, same campaigns, same console. Choose it when footprint is cheap and you want everything local; choose the C agent when you want on-demand modules, or elevation, which the C agent alone offers. Both run in one fleet, side by side.

New check-ins land in a pending queue, not in your fleet. An operator with the right grant admits an agent to a campaign — or rejects it. Rejection is type-to-confirm and final: the agent is ordered to destroy itself, and the decision is logged like everything else. Stale and one-way hosts prune cleanly.

Operators sign in with a password and a TOTP second factor — nothing to install, nothing to license per machine, nothing to carry. An operator's workstation stops being an installation and becomes a credential. Multi-operator is the default way the product exists: every operator sees the same fleet at the same time.

The panel ships fully localized in English and Russian — every view, every dialog, every audit string. Language is a per-operator setting, not a per-server one.

For teams that want native windows and tray alerts, the desktop client is a full native reimplementation of the panel — not a wrapper around a web view. It wins nothing for existing: if a capability cannot be reached from a browser tab, it does not ship. Research note N.01 is the doctrine; the console page is the proof.

A module fetch is a one-time download token, a per-module key derived from the session, results streamed back in chunks, and key material wiped on unload. The server gates what an agent may fetch by agent type and platform. And the fetch never names itself: capable agents pull each module under a random cover name with a matching content type, so the request reads as an ordinary asset — a font, a script, a map file — not as dpapi.dll. A quiet agent therefore carries almost nothing — the loud moment is the fetch, and it happens once, in memory, wearing a boring name.

Where there is no loader — the static embedded builds — the same module set is compiled in. One codebase, two linking decisions, no fork in capability. The budget table and the full fetch sequence are in research note N.04.

The builder takes a traffic profile per build, so one campaign can read like cloud telemetry and the next like developer tooling — from the same teamserver. Profiles are part of the licensed material and ship ready to use, not as templates you finish yourself.

An optional Cloudflare Worker fronts the teamserver so agents never see its address. Server keys rotate without a fleet rebuild. And when the path home dies entirely, Ghost Chain recovers tasking through a blockchain resolver — a fallback we would rather you never need, built for the day you do.

A stager URL is a token with a use count. Served once, it burns; burned, it answers like nothing was ever there. Payload pools go further — one URL serves a different build per delivery, so a recovered sample describes one artifact, never your fleet.

Different guards call for different doors. The full stager family is the choice for hosts watched by CrowdStrike, SentinelOne, or Kaspersky; the minimal family — smaller, metamorphic, with a fully in-memory variant — is built for Symantec, Defender, and static-ML environments. The matrix of which family covers which product ships in the licensee manual, and it is honest about the cells marked verify first.

Routing is only half of the discipline; the other half is measurement. Every build is exercised in the evaluation lab — static, on-execute, and behavioral lanes across the tracked engine families — and rev 0328 completed the matrix with no blocking verdict. The method is on the home page.

A signature written against one AXIS artifact describes exactly one artifact. The pipeline mutates every binary it emits — polymorphic API hashing, junk and section variation, padding — and the server re-mutates every Windows artifact it serves — timestamp, provenance header, checksum, overlay — so no two downloads share a hash, even of the same payload. The persistence installer goes further: its stub is compiled per install, on the server, with randomized optimization, names, and a 192-combination identity pool for the provider names it registers. Investigators hold samples; they never hold the sample.

The MSI pipeline packages a Ghost Stream install and signs it in the same pass. The builder streams a live log as it works — profile, size, mutations — and keeps an artifact history your whole team can see, not just the operator who pressed build.

BOFs are contained by design: a fault in a tasked tool comes back as an error line in the workspace, and the agent keeps its session. The guard is the feature — the library being large is just nice.

On a host where spawning cmd.exe or powershell.exe is the loudest move you can make, Phantom answers commands without creating a process at all. A 116-command built-in set covers the everyday work in-process; the managed runtime loads lazily, only when a command genuinely needs it — so a session that never leaves the built-ins produces no module-load telemetry at all. Color-coded in the console so you always know which shell you're in.

Execution inherits the token context you task it with, so shells, BOFs, and modules act as the identity you chose — not as whatever the process happened to be.

Mint stagers from CI, sweep results into your own evidence store, export audit on a schedule, wire build completion to whatever chat your team actually reads. If you can reach the panel, you can reach the API — same origin, same credential.

# tonight's build, from cron
curl -sS -X POST "$TS/api/builds" \
  -H "Authorization: Bearer $TOK" \
  -d '{"profile":"windows-x64","traffic":"ms-graph","campaign":"NORTH"}'

# audit trail, to the evidence store
curl -sS "$TS/api/audit?campaign=NORTH&format=csv" \
  -H "Authorization: Bearer $TOK" -o north-audit.csv

Directory discovery exports straight to BloodHound, so collection becomes a graph your team already knows how to read. The audit log exports to CSV for evidence stores and after-action reports.

The object runtime is Beacon-compatible, so the community's tooling — and your own — runs without a port. Peer chaining routes over SMB named pipes for segmented networks; SOCKS5 and reverse tunnels carry the rest of your kit through the agent that is already there.