SECURE REMOTE ACCESS + RMM · BUILT FOR MSPs

Remote access you can prove.

ManagedBit is a security-first remote control and RMM platform. End-to-end encrypted at 60 frames a second. Consent enforced by mechanisms, not policy documents. And honest about every feature — including the ones we haven't shipped yet.

0fps
SUSTAINED · MEASURED 57–61
~0ms
CAPTURE→GLASS P50 · 1440p
0
PLAINTEXT BYTES AT THE RELAY
0ms
REVOKE → SESSION KILLED
0
INDEPENDENT SIGNING-KEY ROLES
0
DESIGN DOCS · ADVERSARIALLY REVIEWED

THE NON-NEGOTIABLES

Six rules nothing ships around.

These aren't values-page decoration — they're standing engineering constraints, and several are enforced by tests that fail the build.

End-to-end encrypted. Always.

The relay forwards ciphertext. The control plane authorizes sessions it can never decrypt. There is no plaintext middle, ever — including every fallback path.

No covert access.

A non-removable, fail-closed presence indicator and a tamper-evident audit chain on every session and every channel. That rule binds us and our own staff too.

Fail closed.

A security control whose precondition is missing refuses — it never warns and proceeds. Stand-ins are held to the same bar: if it can't enforce, it doesn't run.

Never our own crypto.

Vetted primitives only — Noise, SPAKE2, ChaCha20-Poly1305, BLAKE3 — version-pinned, with the risky dependencies named and an external audit budgeted.

Memory-safe by construction.

Agent, viewer, relay, control plane: all Rust. The C surface is a thin, audited boundary at the OS and codec edge — the CVE class that haunts this industry is designed out.

Recovery belongs to the customer.

Anything recoverable recovers with your keys — BYOK, customer KMS, optional quorum escrow. Vendor-held master keys are structurally absent.

CONTROL PLANE
authorizes · signs grants · never holds session keys

SIGNED AUTHORIZATION ONLY

TECHNICIAN
viewer · keys live here
SE / TPM / FIDO2
RELAY
sees: ciphertext · size · timing
holds: zero keys · reads: nothing
ENDPOINT
agent · keys live here
SE / TPM · consent banner

Breach the middle and you hold noise. Keys are agreed endpoint-to-endpoint; the relay refuses unauthorized sessions but can never open authorized ones.

LAYER 01 CRYPTOGRAPHIC FOUNDATION

The relay can't read a pixel. Neither can we.

Every session is encrypted endpoint to endpoint. Everything in between — including our own infrastructure — forwards ciphertext and authorizes sessions it can never open. Keys live in silicon, proofs replace self-reports, and a stolen signing key mints refusals, not access.

NOISE XXpsk3 / IK  ·  ChaCha20-POLY1305  ·  X25519  ·  BLAKE3
SPAKE2 PAKE  ·  P-256 IN SECURE ENCLAVE / TPM  ·  WEBAUTHN

End-to-end encrypted, always

LIVE

Screen, input, terminal, files, clipboard — one authenticated Noise connection, per-channel AEAD, and no plaintext middle on any path.

ENCRYPTION OVERHEAD: NOT MEASURABLE AT P50
  • Ad-hoc sessions run Noise XXpsk3 with the PAKE output as the pre-shared key; unattended sessions run Noise IK against grant-pinned static keys.
  • Media rides QUIC datagrams, channels ride prioritized QUIC streams — all inside one authenticated session, 24 bytes of overhead per datagram.
  • Session identity is address-independent: rekeying, resumption, and every channel survive a network change untouched.
  • E2EE holds through every relay and fallback path — sovereignty survives the firewall.

A zero-knowledge relay

LIVE

The most exposed component holds the least: no keys, no frames, no tenant metadata — and it refuses sessions the control plane didn't sign for.

  • Signed relay reservations carry a session id and expiry — nothing else. The relay can gate abuse without learning who you are.
  • Queue management drops by age, never by content; the relay never inspects ciphertext to decide what matters.
  • Production-hardened: session-sharded workers, drain-mode restarts, live signing-key rotation with no restart, fail-closed keyless boot.
  • Impairment and throttling knobs are compiled out of production relay binaries — a shipping relay physically lacks a degradation code path.

Keys that live in silicon

LIVE

Device and technician identities are held in the Secure Enclave or TPM — usable on that machine, never copyable off it.

  • Hardware is mandatory where present; software keys remain only as the honest fallback on hardware-less machines — and they say so.
  • Registering a hardware key auto-revokes that identity's software keys: no theft-resistance theater while a stealable key stays valid.
  • A fleet re-key ceremony migrates existing software fleets into hardware without re-enrolling — continuity-proven, refused for clone suspects.
  • Retiring a device destroys its hardware-held keys and wipes the credential, with proof in the audit chain.

Proven, not self-reported

LIVE

The control plane cryptographically proves a key lives in genuine hardware — policy sees verified verdicts, never the endpoint's word.

  • The full TPM 2.0 MakeCredential / ActivateCredential ceremony, implemented in pure Rust in-process — self-hostable anywhere.
  • 2,160 manufacturer trust anchors, each classified silicon vs virtual; a VM fleet attests honestly as virtual and can't launder itself up a tier.
  • Verdicts are pure recomputations over stored evidence: import a trust root and the estate re-verifies with zero agent round-trips.
  • Endorsement-key uniqueness doubles as cryptographic clone detection; duplicates lose unattended access until a human clears them.

A stolen key mints refusals

LIVE

Seven separate signing roles behind an HSM-capable seam — and endpoints verify every grant against an independently-signed directory.

  • Mint, directory, brand, presence, inventory, monitor, deploy: one stolen key can never do another's job — message spaces are disjoint and tested.
  • Agents check every session's technician against a hash-chained, independently-signed key directory — verified live: a validly-signed grant from a stolen mint key was refused at the endpoint.
  • Keys rotate under dual old+new signatures with fleet-wide replay convergence — verified mid-session with zero frame interruption, including file↔HSM moves.
  • Private keys can live on a PKCS#11 token — non-extractable, self-hostable, with a cloud-KMS adapter on the same seam.

Compromise has an expiry date

LIVE

Sessions re-key continuously and re-prove both identities on a clock — a captured key or session state goes stale in minutes.

  • Per-direction key epochs ratchet roughly every two minutes with zero coordination messages.
  • A fresh Diffie-Hellman re-handshake runs inside the encrypted session every 10 minutes — invisible to the relay — evicting any passive key holder.
  • Both ends re-sign the transcript at every round: the technician's key and the device's enclave identity.
  • Resume credentials rotate every round; a dropped link heals through a 120-second window with authorization re-checked, in ~0.5 s.

A human touch per session

LIVE

Protected tenants require a live hardware-key touch for every connection — and every resume. Stolen cookies stop being enough.

  • The WebAuthn challenge is bound to the connection's handshake hash — per-connection-fresh by construction, so a captured assertion is inert anywhere else.
  • Nothing opens before the proof: no channels, no capture, no input — refusals are structural, not UI states.
  • An 8-digit approval code derived independently by viewer and console means a racing malware prompt shows the wrong number.
  • A grant requiring proof refuses to run against a viewer too old to perform the ceremony — skew fails loudly, never silently.

Codes that can't be brute-forced

LIVE

Attended support runs a real PAKE: short codes with no offline attack, and a spoken four-word check that defeats even a hostile relay.

  • SPAKE2 (the same construction that protected a decade of browser-based PIN pairing): the only way to test a guess is a live exchange, and one failure burns the code.
  • Codes are single-use with a 60–120 s TTL; the code itself never crosses the wire.
  • Both ends derive the same 6-digit SAS from the handshake; two humans comparing it makes interception visible — even by the relay.
  • The viewer gates input and video behind the SAS confirmation; unattended mode replaces ceremony with pinned-key cryptography, never with "trust us."

LAYER 02 CONSENT & TRANSPARENCY

No covert access. Structurally.

The person at the machine always knows — and their consent is enforced by mechanisms, not settings. If the indicator can't be proven visible, the session dies. If a notice can't render, the read refuses. And none of it can be bought away by a pricing tier.

VISIBILITY-PROOF WATCHDOG  ·  PER-DISPLAY CHIPS  ·  KNOCK
CAPTURE-EXCLUDED PROMPTS  ·  HASH-CHAINED AUDIT

The banner that can't lie

LIVE

Screen sharing always shows a visible indicator — and visibility is proven, not assumed. Cover it, minimize it, wedge it: the session ends.

OCCLUSION → AUDITED TEARDOWN: ~100 ms
  • Occlusion, minimize, other-desktop, and display-sleep are all judged per tick; any uncertainty fails closed — no code path returns "visible" on error.
  • The privileged service challenges the banner process on its own clock with single-use nonces — a frozen, suspended, or replaced UI cannot answer, so it cannot keep a session alive.
  • Every captured monitor carries its own load-bearing presence chip; a covered chip ends the session naming the display.
  • The user can move and shrink the indicator — but never below a perceptibility floor, and remote input is latched out of dragging it.
  • The badge is exact truth: SCREEN ×N counts live captures; a terminal-only session never claims "screen shared."

Ask the person at the machine

LIVE

On "knock" devices a connection waits at the door — nothing captures, nothing injects — until a real human clicks Allow on a prompt built to resist fakery.

  • The prompt is jittered per invocation, requires a settle delay, has no keyboard shortcut to Allow, and the click only counts if the prompt was provably unoccluded at that instant — decoy overlays lose.
  • It names the technician from the signed key directory — never free text an attacker could set to "Windows Update."
  • Deny and timeout are terminal, audited, and arm a per-technician cool-down against prompt-fatigue attacks; the person at the machine can clear it.
  • Long sessions re-ask at every re-authorization: "approve each session" can never silently become "approve forever."

Our own UAC

LIVE · macOS FIRST

Sensitive actions mid-session prompt the person at the machine — on a surface the technician can neither see nor click.

  • The prompt is excluded from the outbound stream as a per-frame admission gate — pixel-proven absent from the viewer in ~78 ms, not hoped absent.
  • Injected input near a prompt is latched out; the remote hand cannot click Allow on its own request.
  • Standing consents are born and stored at the machine with an anti-rollback epoch — the control plane mirrors them and can never author one.
  • Headless machines can be pre-authorized by an admin, but never while a human is provably present; unknown means ask.

Watching the watchers

LIVE

Monitoring the monitors is a product feature: clients see what their MSP watches, and monitoring-shaped behavior gets flagged automatically.

  • A sustained-polling detector flags legitimate one-off reads that turn into a surveillance stream — aggregated across technicians and devices so spreading the load can't evade it — and the flag lands on the client's audit surface.
  • Every client org has a transparency card listing every monitor, event rule, rollup, and remediation binding imposed from above it in the tree.
  • When a technician reads live human activity (processes, logged-in users), a named attribution toast renders at the machine first — and if it can't render, the read refuses.
  • Privacy-forbid-wins: an MSP can enforce operational monitoring down the tree, but a client's privacy refusal beats it — proven at both ends.

An audit chain, not an audit log

LIVE

Hash-chained, signed, append-only records on both the endpoint and the control plane. Edit, reorder, or delete anything and verification fails loudly.

LAST FULL VERIFY: 8,211 ROWS · 0 BROKEN LINKS
  • A failed audit append is session-fatal — an unaudited session does not exist as a state.
  • Content never enters the chain: lifecycle, counts, hashes, and resolved paths — never keystrokes, file bytes, or clipboard.
  • Audit outlives tenant deletion by design: operational data cascades, the record does not.
  • Mid-session network moves are recorded — "this session hopped to an unknown network" is exactly what incident reconstruction needs.

The OS indicator tells the truth

LIVE

Screen capture exists only while a screen session does — an idle managed machine is never "being recorded," and the OS's own indicator agrees.

  • Capture is constructed at session establishment and stood down when the session-end event passes the audit chokepoint — every end path, by construction.
  • Terminal-only and files-only grants construct no capture at all and stream zero media — verified, not asserted.
  • Lock, fast-user-switch, or a credential prompt interrupts in under a second; resume must re-bind to the same OS session and user or it refuses.
  • An idle machine shows no banner at all — the indicator appears for things worth knowing, so it never trains people to ignore it.

Consent posture is signed

LIVE

Silent, notify, knock, or deny-while-in-use — per device, signed under its own control key, so posture can only ever tighten out from under an attacker.

  • Presence policy travels as a device-bound, strictly monotonic signed document — a stolen mint key cannot author consent posture, and rollback is fork-evident.
  • Every session delivery carries the mint-time posture as an enforcement floor; the agent enforces the stricter of the two.
  • Enroll tokens set a default posture at birth, so BYOD and server classes start life with the right rules.
  • The tray shows "Managed by your MSP" and lets the machine's owner see — and revoke — standing consents.

The banner survives the pricing page

OWNER DECISION

No tier, add-on, or flag removes the consent indicator. We deliberately forfeited the covert-monitoring market so "no banner" is never ambiguous on any endpoint.

  • The buyer is not the protected party: the MSP pays, the person at the screen is protected — the payer must not be able to remove a third party's protection.
  • Tiers may sell form, never presence: whitelabel styling, a compact pill, repositioning — the indicator itself is constant.
  • Where we can't prove the indicator is visible (some Linux desktop stacks), we refuse to host the session rather than ship a covert-capture hole.
  • The direction of travel is more visibility, not less: approval flows, end-user notifications, audit reporting.

LAYER 03 THE LIVE SESSION

Fast enough to forget it's remote.

A screen-content codec that keeps text bit-exact, loss recovery measured in kilobytes, a rate controller that survives hostile networks — and every capability as a separately-permissioned channel inside one encrypted connection.

57–61 FPS MEASURED  ·  LOSSLESS TEXT OVERLAY  ·  ~3 KB LOSS HEALS
ROAMS IN ~5 ms  ·  TERMINAL · FILES · CLIPBOARD · MULTI-MONITOR

60 fps, ~20 ms, encrypted

LIVE

Hardware HEVC capture-to-glass with encryption on — measured, on both desktop platforms, at desktop resolutions.

2560×1440 · 57–61 FPS · P50 ~20–30 ms · 0 TILE ERRORS
  • Zero-copy GPU pipelines on both platforms: ScreenCaptureKit → VideoToolbox and DXGI → Media Foundation, with damage tracking so static pixels cost nothing.
  • HDR and wide-gamut desktops tone-map correctly instead of washing out — the classic remote-desktop tell, handled at capture.
  • An opt-in GPU-resident render path keeps decoded frames on the GPU end-to-end, honestly measured per scene type.
  • Engineered toward multi-monitor 4K60; every claim on this page is a measured number, not a target.

Text is never blurry

LIVE

A content classifier routes every tile: text rides a lossless overlay above the video stream, bit-exact — even while scrolling.

EDGE FIDELITY 1.0 VS 0.88–0.95 VIDEO-ONLY · −59% BANDWIDTH MASKED
  • Every 128×128 tile is hashed, deduplicated (~44× fewer tiles per frame), and classified — text to a lossless palette codec, imagery to HEVC, solids to a fill command.
  • The video encoder skips regions the lossless layer owns: −59% bandwidth on text-heavy screens at identical fps and latency.
  • Verified scrolls ship as 13-byte move records plus only the changed boundary rows — text stays pixel-crisp during the scroll, and a copy is emitted only for pixels proven byte-identical.
  • Keyframes are demand-only: no periodic bitrate spikes, ~1 Mbps saved at idle, and recovery never depends on a lucky GOP boundary.

Loss costs kilobytes, not seconds

LIVE

Packet loss heals through tiny reference refreshes and per-frame parity — a momentary softness, not a frozen screen waiting on a keyframe.

TYPICAL HEAL: 2–4 KB · ~200× SMALLER THAN A KEYFRAME
  • Long-term-reference refreshes repair loss in ~30 ms at steady-state latency; full keyframes are the last resort, not the first.
  • Every frame carries XOR parity — single lost packets reconstruct locally, with no round trip; large sync frames carry interleaved parity groups so bursts heal too.
  • A decoder output-vacuum net catches the unfixable case and demands a genuine keyframe in ~0.5 s — freezes are bounded by design.
  • Soak-tested continuously: 1-in-300 loss for a minute → every drop healed, zero divergence, frame rate held.

Built for bad networks

LIVE

The stream measures its path and degrades in a deliberate order — then climbs back on its own. On a clean network, all of it costs exactly nothing.

OVERSHOOT → CONVERGED IN ~0.7 s · VERIFIED ON ~200 ms CROSS-CONTINENT LINKS
  • A delay-gradient controller cuts once to measured capacity — never a spiral of guesses; probe clusters rediscover headroom and lift the target back.
  • The degradation ladder is explicit: bitrate → overlay budget → frame rate → resolution (½, ¼) → software quality — and it climbs back in reverse.
  • The technician chooses crisp text vs smooth motion; a constrained session says so in the status bar instead of leaving you guessing.
  • A standing parity gate re-verifies after every change that a clean path is bit-for-bit the fast path.
  • Bulk file transfers yield to the picture from end-to-end evidence — and cost the transfer nothing on a fat pipe.

Sessions that refuse to die

LIVE

Networks change, processes crash, machines reboot — the session heals through all of it, and revocation still bites mid-heal.

ROAM ~5 ms · REBIND 130–500 ms · CRASH-RESUME ~0.5 s
  • Walk from Wi-Fi to a hotspot mid-session: active roams heal in ~5 ms; NAT rebinds in 130–500 ms; OS network monitors on all three platforms trigger the move instantly.
  • Past the roam budget, a 120-second resume window catches everything — cryptographically, with authorization re-checked and denial terminal.
  • Kill the viewer process: it resumes from a fresh process in ~0.5 s. Kill the host: it boots straight into "interrupted" off a keystore-held arm and heals.
  • Every path was soak-tested together: roams × rekeying × forward-secrecy rounds × loss, zero interruptions.

Every monitor, your layout

LIVE

Each remote monitor is an independent stream — one focused at full rate, the rest as live thumbnails that pause themselves under pressure.

  • Selector strip, remote-arrangement grid, break-out windows, and pin-to-local-display — exact across mixed-DPI monitors, with per-device layout profiles that restore on reconnect.
  • Previews run at 2 fps / ~300 kbps and auto-pause under bandwidth pressure with an honest "paused for bandwidth" label.
  • Input carries its target monitor and injects only onto currently-captured displays — an ambiguous event is dropped and counted, never guessed.
  • Hot-plug, sleep, and wake re-aim capture by stable hardware identity in about a second; follow-active-monitor tracks where the user is actually working in ~6 ms.

More than a screen

LIVE

Terminal, files, and clipboard are first-class channels with their own permissions — usable together or entirely without video.

  • Terminal: real PTYs (up to 8), full TUI support, environment scrubbed of agent secrets, bytes never logged, children reaped on every teardown path — revocation ends the effects, not just the transport.
  • Files: resumable transfers with whole-file BLAKE3 verification and atomic finalize; read and write are separately grantable — the DLP toggle is one policy line.
  • Clipboard: the technician's clipboard crosses only at an explicit paste gesture; the endpoint's pushes on change — each direction separately licensed, direction-badged on the banner.
  • Terminal-only and files-only sessions stream zero media and construct no capture; the session model matches the consent given.

Works on a potato

LIVE

GPU-less VMs and headless servers stream through a bespoke software codec with no video decoder at all — the case that breaks tools derived from game streaming.

REAL GPU-LESS VM · 1080p · 19–32 FPS · ZERO DECODER CONSTRUCTED
  • When no hardware encoder exists, tiles become the stream: intra-only, per-tile, loss-resilient, decoding straight onto a persistent canvas.
  • Degradation rule: drop fps, raise photo quantization — never blur text, in software mode too.
  • Engaged organically and honestly: a machine with any real encoder rides the hardware path; forcing is a dev knob, not a guess.
  • The adaptive controller has a software-mode rung, fail-closed behind a platform-adoption gate.

LAYER 04 THE CONTROL PLANE

Multi-tenant to the bone.

An MSP-shaped tenant tree with database-enforced isolation, policy-as-code that fails closed, grants that expire and re-authorize themselves, and enrollment that treats every new device as unproven until a human says otherwise.

POSTGRES RLS DAY 1  ·  CEDAR POLICY · DENY-BY-DEFAULT
REVOKE→KILL ~18 ms  ·  OIDC + WEBAUTHN STEP-UP

Tenancy that matches reality

LIVE

MSP → client → site, with umbrella groups at any depth for resellers and holding companies. Acquisitions are a subtree move, not a migration project.

  • Tenant move is cycle-safe and first-class; policy inheritance follows automatically because the hierarchy is rebuilt from truth on every request.
  • Deletes tombstone and cascade the subtree — devices, tokens, grants — while audit outlives everyone, by design.
  • Isolation is enforced by the database: forced row-level security from the first migration, a no-bypass role, per-request visible-tenant sets.
  • Escape hatches are two narrow, audited functions — and the codebase carries a standing rule against the query shape that would fail open.

Policy as code, deny by default

LIVE

Tenants author real authorization policy — validated at write, inherited down the tree, with parent forbids that children cannot override.

  • Powerful capabilities are deny-by-default: unattended access, scripts, file directions, clipboard directions — each an explicit per-tenant decision, conditionable on device class.
  • The engine denies on any policy error — an erroring forbid can never be silently skipped — and evaluation loads only the resource's ancestor chain, property-tested.
  • Dry-run a draft against a live device; the fleet-impact preview counts denied devices and live sessions a change would end, with a save-time confirm.
  • When a permission splits into finer ones, stored tenant policies are rewritten, re-validated, and audited at boot — never silently voided.

Access that expires by itself

LIVE

Access is a short-lived signed capability, re-authorized on a schedule against current policy — and revocable to a dead session in tens of milliseconds.

REVOKE → ENDPOINT KILL: ~18–35 ms MEASURED
  • Grants are device-bound, capability-sliced, and independently enforced by the endpoint — the same list the policy engine decided, checked twice.
  • Long sessions re-run the entire mint authorization as the original technician on a schedule; offboard someone and their live session dies at its next bound.
  • Revocation bites at launch, in-session, and at resume — with reconnect snapshots converging agents that were offline.
  • Concurrent-session caps per technician, tenant, and device evaluate race-safe as policy context — the fan-out signature of a stolen key hits a wall.

Enrollment that can't be rushed

LIVE

Devices are born pending. A leaked installer token mints nothing — it buys a place in an approval queue, in front of a human, with risk flags attached.

  • Enroll tokens carry their own guardrails: CIDR allowlists, hourly velocity caps, use limits, expiry — and auto-approve refuses to exist without an IP allowlist.
  • Golden-image clones and re-imaged machines announce themselves — key × hardware-fingerprint disagreement plus endorsement-key uniqueness — and suspects lose unattended access until cleared.
  • Fingerprints are board-anchored signals, never authentication; a swapped disk degrades gracefully instead of forking identity.
  • Retire is a real ceremony: credentials wiped both sides, hardware keys destroyed, grants revoked, proof in the chain.

Identity your way

LIVE

A bundled Apache-2.0 identity provider behind a backend-for-frontend — tokens never touch the browser, and technicians start with exactly nothing.

  • First login JIT-provisions a technician with zero access until an admin grants a subtree; grants, disables, and revokes are all subtree-contained.
  • Teams make forty technicians one grant; service accounts give integrations their own read-only principal with per-token IP allowlists; an auditor role covers compliance seats.
  • Dangerous verbs demand a fresh hardware-key touch at the moment of the action — not a ten-minute window a stolen cookie can ride.
  • One-verb offboarding: login refused, console sessions dead, grants revoked, mints refused, live sessions ended at their next bound.

Lock the doors, keep your keys

LIVE

Pin your whole environment to your networks — and recover from your own mistakes without the vendor ever being the recovery path.

  • Network login policy makes stolen credentials unusable off your CIDR ranges — enforced across every authority path: visibility, roles, and containment.
  • The editor computes who a change would evict — including your parent MSP's admins — and refuses accidental self-lockout.
  • Break-glass accounts are local, WebAuthn-only, IP-pinned, verb-restricted — live-verified under a real lockdown; self-hosted deployments have no vendor to call, and don't need one.
  • Security notifications are written in the same transaction as their event; per-tenant lanes mean one tenant's storm can't bury another's alarm.

A push plane with receipts

LIVE

Every endpoint holds one outbound connection carrying grants, revocations, policy, inventory, alerts, and jobs — with freshness measured by evidence, not socket optimism.

  • Liveness is receive evidence: pings every 5 s, a 16-second silence deadline, and network changes tear down and redial immediately — a half-open socket can't fake presence.
  • Unattended sessions carry a fail-closed lease: control plane unreachable past 60 s means the session ends — revocation enforcement outranks availability.
  • Signed documents converge on reconnect snapshots; imperative one-shot actions are never re-pushed — the double-actuation stopper.
  • The console's "live right now" is mirrored from the endpoint's own audit chain, with phantom rows reconciled and stale rows swept with proof.

Whitelabel without a spoofing surface

LIVE

Your logo and colors on the console, the client, and the consent banner — cryptographically bound to your tenant, and impossible to weaponize.

  • Brand packs are signed under a dedicated key, tenant-bound inside the signature, strictly versioned, and expiring — replay, cross-tenant, and rollback all refuse.
  • Status colors have no field in the schema — "the live glow can't be re-skinned" is structural, not a convention; consent wording and geometry are equally out of reach.
  • Packs are declarative data only — never code — bounded and validated identically at upload and at load by the same function.
  • Every failure lands on the default brand, never a stale one; un-branding is deliberately the fail-safe direction.

LAYER 05 THE RMM

Management with receipts.

Inventory in seconds, monitoring at the edge, self-healing with hard floors, and fleet scripting that takes three independent keys to run — all built on the same consent and privacy machinery as the session core.

11 INVENTORY SECTIONS × 3 OSes  ·  EDGE-EVALUATED MONITORS
EXACTLY-ONCE ACTIONS  ·  SIGNED PRIVACY FLOORS

Inventory in seconds, not scan windows

LIVE

Software, patches, services, hardware, posture, network, certificates, users — from one agent on Windows, macOS, and Linux, kept fresh by deltas.

STEADY STATE: A HANDFUL OF HASHES PER DEVICE PER DAY
  • Content-addressed sync: the endpoint offers hashes, the server asks only for what changed — an install surfaces in seconds, not at the next scan window.
  • "Which machines run this app below version X" is an indexed join with per-ecosystem version ordering, not a text scan.
  • The change journal answers when things changed — fleet-wide app updates, the day disk encryption went off — with honest uncertainty intervals.
  • Freshness is stamped by the server's clock; a compromised endpoint can't stamp fabricated data "now."

Privacy-engineered collection

LIVE

Sensitive collection is off until a signed tenant document turns it on — and some things are structurally not collectible at all.

  • The privacy floor is a signed document under its own control key — not a server setting, and not reachable by anyone who compromises the server.
  • PII is tiered per field, with shorter retention clocks that purge on their own schedule; tenants can shorten further, and tighten-only knobs can never switch anything on.
  • Structural non-goals, kept by construction: no browser history, no keystrokes, no screenshots outside a consented session, no file-content indexing.
  • A failing test guarantees self-reported inventory can never silently become an access-control input — the classic RMM fail-open, designed out.

Monitoring at the edge

LIVE · macOS FIRST

Monitors evaluate on the endpoint against a signed per-device policy — the wire carries state changes, so cost scales with problems, not fleet size.

  • Sustain timers persist through reboots — proven live: a condition raised on schedule across a host kill and restart.
  • Alerts ride a durable acked queue: a dropped packet can't lose one, and a crash-looping machine can't use its crashes to suppress one.
  • An authoritative reconciler means the console can never show a stale "firing" — and a manually-resolved alert whose condition persists comes back.
  • The control plane watches the watchers' absence too: agent-offline and inventory-stale alerts with boot grace, so a restart can't false-fire fleet-wide.

Events with a structural conscience

LIVE · macOS FIRST

OS event capture where the privacy class of every source is decided by shipped code — not by whoever wrote the rule.

  • Rules can only reference — and only narrow — a catalog compiled into both binaries; no authored field can widen what's readable.
  • Human-activity and content classes require two keys to enable (the monitor policy and the signed privacy floor) — a single stolen key taps nothing.
  • A reserved security lane means a noisy pack can never crowd out tamper events; budgets degrade loudly, never silently.
  • Secrets that land in log lines are redacted on the endpoint before they reach storage; retention purges PII text on a shorter clock than the events.

Self-healing with a floor

LIVE · macOS FIRST

A crashed service restarts itself at 3 a.m. — inside limits stronger than the ones humans get, with a tamper-evident attempt trail.

  • Remediation authority is content-pinned and expiring: the action, its scope, and its definition are hashed together — drift suspends it.
  • Protected-service floors (repo constants plus signed tenant additions) mean a monitoring key can never relax target protection.
  • A presence floor keeps self-healing off machines your policy says not to touch while someone's using them — deferrals are journaled, once.
  • Monitor packs are shareable, versioned, and structurally incapable of carrying code; adoption previews show per-descendant privacy impact before you commit.

See it now — never stored

LIVE

Live CPU, memory, disk, and network at 1 Hz in the browser; processes and logged-in users on demand — audited, TTL-cached, structurally never historized.

  • Live watches are bounded 15-minute windows with in-stream re-authorization — a disabled technician's stream dies mid-watch.
  • Human-activity reads render a named notice at the machine first; five-minute rollup history keeps the graphs without a firehose, with honest gaps instead of invented lines.
  • Watch-minutes feed the surveillance detector — per device, so one workstation watched all day can't hide in a subtree average.
  • Session-less micro-actions (restart service, kill process) take a hardware touch, show a notice, record before/after evidence, and execute exactly once even across a crash.

Running code takes three keys

LIVE · macOS FIRST

Fleet scripting where the platform's authorization, a signed content directory, and the technician's own hardware touch must all agree — or nothing runs.

  • The technician's WebAuthn touch signs the job, the script hash, the parameters, and the enumerated target set — none of it re-bindable after the fact.
  • Every agent re-verifies all three authorities; a stolen server signing key mints refusals, never execution — proven live against a deleted script.
  • Exactly-once by construction: a crash mid-run finalizes unknown, never failed; results re-serve without re-executing; process trees die together.
  • If someone is at the machine, they see a named "running a script with system privileges" notice first — no notice surface, no run.
  • Parameters pass as argv/env only — no interpolation API even exists — and secret parameters refuse until a real vault ships.

The deployment arc

DESIGNED

Blueprints, desired-state convergence, and patching — designed in depth with the per-platform truths stated honestly, landing phase by phase. The convergence engine is landing now.

  • Declare what a fleet should look like — pin, minimum, tracked, present, absent, blocked — and let devices converge, even after weeks offline. The agent-side engine is already live-verified.
  • Parent-tenant blocks are forbid-wins: "never this software" survives any child's install intent.
  • Patching engages each OS on its own terms — native update APIs, per-distro package managers, and honest report-and-nudge where platforms have closed enforcement to third parties.
  • Rollouts get rings with health-gated promotion — halting on post-install health deltas, not just install exit codes — and vulnerability-driven expedite from live exploitation data.
  • A community marketplace is planned supply-chain-first: transparency logs, pinned adoption, cooldown windows, and a revocation kill switch.

LAYER 06 THE EXPERIENCE

An honest console. A real client.

A complete MSP console that never fakes a feature, and a native technician client where session states are first-class screens. If something isn't wired yet, it says so — the same rule this page follows.

21+ ROUTES  ·  ⌘K PALETTE  ·  NATIVE VIEWER
DEEP-LINK CONNECT  ·  TERMINAL-FREE SETUP

A console that never bluffs

LIVE

Every surface without a backend carries a badge naming exactly what it awaits. Unwired actions disable instead of pretending. Refusals render verbatim.

  • "As of 3 days ago · agent offline" beats a silent stale number; capped lists say "showing most recent N"; absence renders as absence, not as failure.
  • A refused action shows the server's exact reason — a denied unattended mint is never silently downgraded to attended.
  • Derived data is labeled derived, with the evidence in the tooltip; self-reported claims are never styled as verified facts.
  • The approvals badge counts live queues only — preview queues can't inflate it.

Operator speed

LIVE

⌘K to anywhere — including "connect to that machine." Scope the whole console to one client. Save your views. Export what you see.

  • The tenant scope lens filters every page — and states plainly that it's a workflow filter while row-level security remains the boundary.
  • Grids do real ops work: faceted filters, grouping, saved views per user, CSV and print export, bulk actions with per-row outcomes reported verbatim.
  • Bulk verbs run in a background jobs drawer that survives navigation; destructive verbs keep their typed-count confirms.
  • Device pages go deep: identity, attestation, all eleven inventory sections, monitors, events, sessions, and a merged timeline — every leaf deep-linkable.

A native client, not a browser tab

LIVE

Session states are first-class screens: the verification ceremony gates input and video, interruptions show a live countdown, endings tell the truth.

  • Seven ending classes with headline, body, and verbatim detail — and a build-failing coverage test so no new error ships unclassified. "Session ended by policy" renders as the system working, not an error.
  • Modifier-aware input translation: chords, Cmd↔Ctrl remap, and release-all when the gate closes so the remote never keeps a stuck key.
  • Terminal tabs pop out into real OS windows and dock back; the dual-pane files workspace does drag-and-drop both directions with resumable transfers.
  • Fullscreen keeps an escape hatch — a hover pill excluded from input forwarding, so clicking End never also clicks the remote desktop.

Click to connect

LIVE

From console to live session in one click — through a handoff designed so a malicious link can never add trust.

  • Launch tickets are minted atomically with their grant, stored as hashes, and verified-then-burned — a failed proof doesn't consume the ticket, a replay is a uniform refusal.
  • The viewer redeems only against a locally pinned origin; a deep link can never seed the trust set — pinning is a deliberate human act, with the TLS identity shown at the decision.
  • A confirm screen renders device, tenant, mode, and capabilities from the signed grant before anything connects.
  • Refused, replayed, and expired paths were all live-verified — each lands on an honest, specific ending.

Setup without a terminal

LIVE

A first-run wizard takes a technician's machine from blank to operational — pin the control plane, generate a hardware-backed key, register it — across as many consoles as they work in.

  • One key per machine, registered per console — each control plane keeps its own row and revocation state, probed concurrently and rendered distinctly.
  • The console distinguishes "a grant was minted against this key" from "a machine recently proved it still holds it" — dead key files can't hide behind fresh-looking timestamps.
  • Unreachable renders as "could not be asked," never as "unregistered" — the wizard never converts a network error into a trust claim.
  • Next: browser-approval registration (approve in the console you're already signed into — no ticket to steal) and passkey bootstrap in the same wizard.

Your brand, everywhere it's safe

LIVE

Console, technician client, and the endpoint's consent surfaces re-skin to your tenant at runtime — inside hard limits that keep branding from becoming phishing.

  • One signed pack re-skins the web console with zero rebuild and the native surfaces through the same validated schema.
  • The decision surfaces stay ours: no logo on the consent prompt, frozen attribution wording, and a reserved strip the countdown and buttons always own.
  • A living styleguide is the QA harness — every component lands there in the same change, web and native both.
  • Runtime brand changes land live and kill switches land within one push — verified end to end.

Help for machines you don't manage

DESIGNED

A no-install, nothing-left-behind support app: the user runs it, reads a code aloud, and gets help — engineered assuming criminals will try to abuse it.

  • The brokering is already live: device-less sessions under full tenant policy, with hard-wired velocity floors that configuration can only lower.
  • The verification ceremony can never be auto-confirmed — the knob is compiled out of the product binary, asserted at build time.
  • Branded support pages on a dedicated domain, an embeddable zero-authority launcher, and first-contact anti-scam friction that's on by default.
  • No enrollment, no persistence, no conversion hooks: when the app closes, nothing remains — and its audit trail survives on the control plane anyway.

Recording, when it lands, stays sealed

DESIGNED

Session recording as an authorized member of the encryption group — compliance-grade, higher quality than the live view, and still unreadable by the vendor.

  • The recorder is an explicit recipient inside the E2EE boundary — never a tap in the middle; its presence shows on the consent surface.
  • Spool now, build later: recordings come out gap-free and seekable because the recorder can wait for every packet — the live session's latency never pays for it.
  • Storage is ciphertext under your organization's key — in your bucket if you want; playback is an authorization event, audited.
  • Offline enrichment: chapters, thumbnails, searchable text, redaction — all after the fact, none of it in the live path.

THE PRACTICE

Attacked before it's built.

Process is a feature. Every security-bearing design goes through independent adversarial review before code exists; every landing publishes a verification matrix run on real hardware; every failure ever caught becomes a permanent regression scenario.

Adversarial review, pre-build

Three independent lenses — hostile peer, privacy, protocol — attack each design before code. Findings are binding. Criticals get found on paper, where they're cheap.

Live verification

Nothing is "done" until a published matrix passes on real hardware — refusal paths included — with the audit chain re-verified at every landing.

A memory of every failure

Roughly fifty live-caught media failures are catalogued with mechanism and fix, and replayed by a deterministic harness so none of them can return quietly.

Tested across an ocean

A standing cross-continent staging path (~200 ms round trip) through a public cloud relay — judged on experience metrics: time at zero fps, freeze duration, glass age.

Fail-closed doctrine

The one recurring bug class in this industry is a control that fails open when its precondition is missing. Here that shape is a named rule, hunted in review, with exceptions argued in writing.

Published posture

The threat model states its non-goals. The security review records what's still open next to what's fixed. Credibility comes from the list of things we say we don't do yet.

"The crypto core is the strongest part of the codebase… nearly every serious finding is one shape — an authorization control that fails open when its precondition is absent."

PLATFORMS & SOVEREIGNTY

Your metal. Your keys. Your jurisdiction.

Both desktop platforms are complete, Linux is deep into its arc — and the whole plane is built to run where you decide, with the honest state of every column below.

CAPABILITYmacOSWINDOWSLINUX
Remote controlX11
Terminal · files · clipboard
Hardware-bound identity
Key attestationMDA ARC
Inventory (11 sections)
Monitoring + self-healing
Fleet scripting
Privileged service topology
Technician viewer

● LIVE    ◐ PARTIAL — REMAINING HALVES TRACKED PER-DOC    ○ PLANNED
WHERE A LINUX DESKTOP CAN'T PROVE THE CONSENT BANNER IS VISIBLE, WE REFUSE TO HOST — BY DESIGN.

Self-host everything

One compose file stands up the control plane; the relay is a single binary on your own node. No PaaS dependencies, no Node runtime, no CDN or font callouts — the console is served by the Rust binary itself.

Keys in your HSM

Control-plane signing runs behind a PKCS#11 seam — non-extractable on your token — with a cloud-KMS adapter on the same seam. Seven roles, independently rotatable, live.

Your data's jurisdiction

Relay and control plane deploy where you choose. Customer-held keys (BYOK) and per-tenant encryption are designed arcs on the roadmap — stated as roadmap, not implied as shipped.

An IdP you can bundle

The bundled identity provider is Apache-2.0 — chosen for license, not fashion — and swappable behind one OIDC issuer URL. Per-client bring-your-own-IdP federation is the designed next step.

ROADMAP

What's next — labeled like everything else.

Preview surfaces exist and say so in the product. Designed means an adversarially-reviewed design is ratified and waiting on build. Planned means we've named it and sequenced it — nothing more.

IN PREVIEW

  • Patch consoleDeadline-relative compliance, approvals, freeze windows — badged until the engine lands.
  • Deployment consoleCatalog, blueprints, ring rollouts.
  • Reports galleryPatch, session/SLA, audit export, executive summary — grid exports already work today.
  • GPU-resident renderingOpt-in on both platforms, honestly measured per scene.
  • Org settings & integrationsPlan, seats, entitlements, PSA hooks.

DESIGN COMPLETE

  • Quick-support suiteNo-install app, branded support domains, embeddable launcher, anti-scam friction.
  • Session recordingZero-knowledge storage, audited playback, offline enrichment.
  • Blueprints & desired stateDeclarative convergence with forbid-wins blocks.
  • Patch enginePer-OS honesty: native APIs, distro managers, report-and-nudge where platforms closed the door.
  • Multi-technician collaborationEscalation, handoff, one-driver enforcement, scoped annotations.
  • Installers & auto-updateSupply-chain-grade: publisher pinning, A/B rollback, anti-downgrade, rings.
  • BYOK / customer-held keysYour KMS as the recovery root.

ON THE ROAD

  • Reachability ladderPort-443 shaping, proxy support, TCP fallback — ranked first.
  • Direct peer-to-peerRace direct against relay and take the winner.
  • Secure-desktop arcUAC, lock screen, login window — with its own mandated review.
  • Browser & mobile viewersSame channel model, honest trade-offs.
  • Marketplace + vendor catalogSigned community content with transparency logs.
  • PAM & secrets vaultCredential injection without exposure.
  • Apple device attestationThe Mac half of hardware proof.
  • Multi-region relaysMeasured nearest-relay selection, failover, IPv6.
  • AI copilotDeliberately last — as another audited, capability-gated participant.

THE HONEST LEDGER

Every feature. Every status. No asterisks.

Generated from the same engineering documents the product is built from. LIVE means landed and verified end-to-end on real hardware. PREVIEW means the surface exists and says so. Nothing here is upgraded for marketing — that's the point.

See it run on your
worst network.

ManagedBit is in active development with early MSP partners. If provable security and honest engineering are how you want your remote stack built — let's talk.

Back to the ledger

Request early access

A human reads every request. Tell us enough to route it.