Skip to main content

Remote Access: RDP

ServiceRadar desktop/RDP access uses the same agent-routed remote-access plane as SSH, but it is a separate graphical workflow. RDP sessions are routed through a selected edge agent, use a dedicated desktop media path for screen updates, and keep target credentials out of browser storage and persisted session metadata.

Experimental — not generally available

RDP remote access is an experimental feature and is not yet generally available. It remains experimental until the reviewed IronRDP connector backend is linked and the helper readiness probe reports connector_ready: true. Base agent installs do not include the RDP helper. RDP-capable artifacts are optional and are hidden from EdgeOps unless the deployment explicitly enables desktop RDP.

Connection Path​

browser RDP renderer
  -> web-ng WebRTC signaling
  -> core-elx desktop media manager
  -> agent-gateway desktop media service
  -> selected edge agent
  -> per-session serviceradar-rdp-adapter helper
  -> registered Windows RDP target

Use the generic control path for lifecycle and low-rate controls such as open, close, keyboard, pointer, resize, focus, quality, and revocation. Use the dedicated desktop media stream for screen updates, cursor updates, frame metadata, heartbeat, close, and backpressure acknowledgements. Do not route production RDP screen traffic through terminal byte frames.

Enablement Checklist​

Before enabling RDP in an environment:

  • Register only trusted RDP targets. Users must not be able to supply arbitrary upstream hosts or ports at session launch.
  • Assign each target to the agent or allowed agent set that can reach TCP 3389 or the target-specific RDP port.
  • Enable the web-ng/core feature flag only in deployments that should expose RDP UI and WebRTC signaling:
SERVICERADAR_REMOTE_ACCESS_DESKTOP_RDP_ENABLED=true
  • Deploy an RDP-capable signed agent artifact only to the agents that should host desktop sessions. Standard base agents should remain the default for the rest of the fleet.
  • Verify the helper locally on the agent host:
serviceradar-rdp-adapter --capabilities

The helper must report schema serviceradar.rdp.helper.capabilities.v1, protocol rdp, ironrdp_backend_linked: true, and connector_ready: true before the agent may advertise remote_access.rdp. While connector_ready is false, the helper also reports connector_ready_reason so operators can distinguish an unlinked backend from an IronRDP-linked helper that still lacks live authentication, media, cleanup, and demo proof.

Target Policy​

An RDP target policy is compiled by the control plane and sent to the selected agent as a trusted per-session snapshot. The browser never chooses the upstream address, credential secret, route, TLS policy, or redirection policy.

Example policy shape:

{
  "target_id": "win-admin-01-rdp",
  "display_name": "win-admin-01",
  "device_uid": "device-win-admin-01",
  "protocol": "rdp",
  "route": {
    "selected_agent_id": "agent-edge-01",
    "selected_gateway_id": "gateway-default",
    "allowed_agent_ids": ["agent-edge-01"]
  },
  "upstream": {
    "host": "10.0.0.10",
    "port": 3389
  },
  "tls": {
    "mode": "verify",
    "ca_bundle_id": "corp-rdp-ca",
    "ca_bundle_pem": "-----BEGIN CERTIFICATE-----\n...\n-----END CERTIFICATE-----",
    "nla_mode": "required",
    "server_name": "win-admin-01.example.com"
  },
  "credential": {
    "mode": "memory_user",
    "allowed_principals": ["EXAMPLE\\\\alice", "[email protected]"]
  },
  "screen": {
    "max_width": 1920,
    "max_height": 1080,
    "frame_rate": 30,
    "bitrate_bps": 8000000,
    "idle_seconds": 900,
    "ttl_seconds": 3600
  },
  "redirection": {
    "clipboard_mode": "disabled",
    "drive": false,
    "printer": false,
    "audio": false,
    "smart_card": false,
    "file_copy": false
  },
  "approval_required": false,
  "recording": {
    "metadata_enabled": true,
    "screen_enabled": false,
    "clipboard_enabled": false,
    "file_enabled": false,
    "audio_enabled": false
  },
  "metadata": {
    "rdp.dial_timeout_ms": "10000",
    "rdp.kdc_timeout_ms": "10000"
  }
}

rdp.dial_timeout_ms and rdp.kdc_timeout_ms are optional trusted target metadata overrides for the helper's TCP dial stage and Kerberos/KDC stage. Keep them unset unless an environment needs a longer stage timeout; the helper caps each value at 30 seconds.

Secure defaults:

  • protocol defaults to rdp.
  • RDP port defaults to 3389.
  • TLS mode defaults to verify.
  • NLA mode defaults to required.
  • Screen recording and clipboard recording are disabled unless explicitly enabled.
  • Metadata recording is enabled when no recording fields are set.
  • Clipboard mode defaults to disabled.

RBAC And Approval​

Grant RDP access separately from SSH:

  • Use devices.remote_access.rdp.open for graphical desktop access.
  • Continue using devices.remote_access.ssh.open for SSH.
  • Keep devices.console.open scoped to console entry points such as Proxmox host shells.

Brokered reusable credentials require approval. This is intentional: brokered credentials are a compatibility escape hatch and must be bound to actor, session, target, route, TTL, and audit metadata before release. Do not use a shared desktop administrator credential as the normal enterprise path.

Credential Modes​

Preferred modes:

  • domain_delegation: the target sees the real enterprise identity through domain-backed delegation when the environment supports it.
  • smart_card: future smart-card or certificate-backed user authentication. Treat this as a separate credential-custody review before broad use.
  • memory_user: the user supplies their own domain credential for one session. The credential is delivered only to the selected agent/helper and dropped on close, timeout, route loss, auth failure, or adapter startup failure.

Fallback mode:

  • brokered_secret: a centrally brokered, target-bound credential. This must require approval, TTL, and target/session/route binding. The agent must receive only the session grant and must not persist reusable secrets locally.

Never store RDP passwords, generated RDP files, smart-card material, or credential caches on agents or in browser storage.

TLS And NLA​

Use NLA-required RDP as the default posture — see the tls block in the Target Policy example above.

TLS modes:

  • verify: verify the target certificate using the configured trust roots and server name.
  • pinned_ca: verify using a named CA bundle managed by ServiceRadar policy.
  • system: use the agent host system trust store.
  • tofu: trust-on-first-use for controlled enrollment only.
  • insecure: lab-only. Do not use in production.

server_name is the TLS certificate identity, not necessarily the dial address. It must be an ASCII DNS name using LDH labels (A-Z, a-z, 0-9, hyphen, and dots). IP literals are rejected for verified TLS. You may dial an RDP target by private IP in upstream.host, but the target certificate must still contain a matching DNS SAN and server_name must use that DNS name.

For production RDP, prefer verify with an explicit ServiceRadar-managed CA bundle or pinned_ca for private Windows/xrdp certificates. The helper path is expected to build a normal Rustls client verifier from PEM CA bundle material, reject DER, empty, or malformed bundles, and disable TLS resumption because CredSSP does not support it. system loads the selected agent host's native trust store and must fail closed if the store cannot be loaded or contains no usable roots. Do not use TOFU as a standing trust model; it implies persistent agent-local state and should stay limited to a reviewed enrollment workflow.

Registered CA bundles are public trust material, not credentials. For registered bundle modes, the trusted session policy should include both ca_bundle_id for auditability and bounded ca_bundle_pem material so the selected helper can verify the target without agent-local trust state. Do not put private keys or passwords in target TLS policy.

Keep nla_mode set to required for Windows RDP. Disabling NLA is a temporary lab-only setting and should not pass production policy review.

Redirection Controls​

All redirection features are disabled by default — see the redirection block in the Target Policy example above.

Clipboard modes:

  • disabled: no clipboard traffic.
  • text_to_remote: browser-to-target text only.
  • text_to_browser: target-to-browser text only.
  • text_bidirectional: text both ways.

Drive, printer, audio, smart-card, and file-copy redirection should remain disabled until each feature has its own policy, audit, malware scanning or export-control posture, and tests. Use ServiceRadar file-transfer workflows for controlled file movement rather than enabling drive redirection by default.

Recording And Audit​

Metadata recording is the default. It records session facts, not desktop content:

  • actor, session, route, selected agent, target, protocol, and credential mode
  • TLS/NLA posture and policy decisions
  • screen resolution, frame rate, bitrate, byte counters, duration, and close reason
  • redirection state and policy decisions
  • approval reviewer metadata when approval is required

Screen frames, screenshots, clipboard content, file contents, audio, and smart-card events are sensitive content. Enable content recording only when an explicit policy, retention period, user-visible session indicator, and export access model are in place.

Example metadata-only recording policy:

{
  "recording": {
    "metadata_enabled": true,
    "screen_enabled": false,
    "clipboard_enabled": false,
    "file_enabled": false,
    "audio_enabled": false
  }
}

Optional Agent Distribution​

The base agent remains the default fleet artifact. RDP-capable releases are optional:

  • The signed release manifest for the base artifact should declare capabilities: ["agent"].
  • RDP-capable bundles should declare remote_access.rdp and include helper metadata, helper protocol version, compatibility range, checksums, signatures, SBOM/license review references, and deployment requirements. Experimental bundles with helper_connector_ready: false must also include helper_connector_ready_reason.
  • EdgeOps hides artifacts with remote_access.rdp unless SERVICERADAR_REMOTE_ACCESS_DESKTOP_RDP_ENABLED=true.
  • One-click rollout installs the RDP helper only when the selected signed artifact declares remote_access.rdp.
  • Installed agents must not advertise remote_access.rdp until the helper readiness probe passes.

For release artifact details, see Agent Release Management.

Operator Workflow​

  1. Enable RDP only in a controlled deployment.
  2. Publish or import a signed RDP-capable agent artifact.
  3. Use EdgeOps to deploy the RDP-capable artifact to the selected edge agent cohort.
  4. Verify the helper readiness probe on those agents.
  5. Register RDP targets with fixed upstream host/port, allowed agent set, TLS/NLA policy, credential mode, screen policy, redirection policy, and recording policy.
  6. Grant devices.remote_access.rdp.open only to the intended roles.
  7. Start with metadata-only recording and all redirection disabled.
  8. Validate one private target before expanding the target set.

Validating A New Deployment​

When bringing up RDP in a new environment, validate against a private target that is reachable from the selected edge agent but not reachable directly from web-ng or the operator browser. A Windows host with NLA enabled is the production-representative path. Register the target as described in the Operator Workflow (step 5), starting with a conservative screen policy such as 1280x720 at 15 fps and raising limits only after backpressure counters stay healthy.

Validation checklist:

  • A user without devices.remote_access.rdp.open cannot list or open the target.
  • A user with devices.remote_access.rdp.open can list only authorized RDP targets.
  • Browser session metadata shows the target, route, credential mode, TLS/NLA posture, recording mode, and redirection state without any credential material.
  • The agent advertises remote_access.rdp only after local config enables RDP and the helper readiness probe passes.
  • Desktop media counters advance only on the dedicated desktop media stream.
  • Browser backpressure updates return credit through the WebRTC control DataChannel and the agent slows or pauses when credit is exhausted.
  • Closing the browser session closes the helper process and clears the memory-user credential grant.
  • Recording/audit views show lifecycle and policy metadata without screen frames or clipboard payloads.

Troubleshooting​

  • RDP action hidden: verify SERVICERADAR_REMOTE_ACCESS_DESKTOP_RDP_ENABLED=true in web-ng/core.
  • Agent does not advertise remote_access.rdp: verify the selected artifact includes the helper, the agent config enables RDP, the helper path is executable, and serviceradar-rdp-adapter --capabilities reports connector_ready: true.
  • Route denied: the target policy selected agent is missing, not in allowed_agent_ids, or not the local agent receiving the open frame.
  • Target connection refused or timed out: the selected edge agent cannot reach the target upstream host and port.
  • TLS verification failed: fix server_name, trust roots, or the target certificate. Do not switch to insecure outside a lab.
  • NLA failure: confirm the Windows host requires NLA and that the credential mode supplies a valid domain/local user for that target.
  • Clipboard denied: confirm the target redirection policy enables the requested clipboard direction.
  • Session closes under load: inspect desktop media backpressure and quota logs for exhausted credit windows, bitrate caps, frame-rate caps, max-chunk rejection, or browser pause/close acknowledgements.