The server is TypeScript on Node/Bun. The client is Go. Operators talk to the server through a web panel or the Electron desktop app, and agents connect over encrypted WebSockets.
After the first start, open `https://localhost:5173`. Default login is `admin` / `admin` unless you set `OVERLORD_USER` / `OVERLORD_PASS`. First startup writes generated secrets to `data/save.json` (inside the container: `/app/data/save.json`) — keep that file private and back it up.
The remote desktop viewer has a **Transport** dropdown with three modes:
- **Canvas** (default): H.264 / JPEG / block frames over the existing WebSocket, decoded into a `<canvas>`. Highest latency, works anywhere the WS does.
- **WebRTC P2P**: browser ↔ agent direct. The server only relays SDP and ICE candidates over the existing WS — MediaMTX is not involved. Lowest latency. Fails when both sides are behind aggressive symmetric NAT.
- **WebRTC Relayed**: agent publishes to a MediaMTX sidecar via WHIP, browser plays via WHEP. The server proxies signaling so the existing JWT auth + per-client RBAC still apply. Lowest-effort fallback when P2P can't punch through.
### Building agents with WebRTC
WebRTC is **opt-in per agent build**. In the builder UI, tick the **WebRTC** checkbox before clicking Build. Without it, the Pion stack (~6 MB of Go modules) is not compiled in — any WebRTC start attempt from the operator returns `webrtc support not compiled in` and the viewer falls back to Canvas. The Canvas path always works regardless of the build setting.
The build tag (`overlord_webrtc`) is also available if you build agents outside the UI:
```bash
go build -tags overlord_webrtc ./cmd/agent
```
### MediaMTX sidecar
Compose starts an `overlord-mediamtx` service for Relayed mode. It needs:
- Port `8189/udp` and `8189/tcp` reachable from operators (WebRTC ICE traffic). The Windows / macOS compose publishes these; Linux uses host networking and shares the host's interfaces directly.
- No auth config — the Overlord server proxies every WHIP/WHEP request through `/api/webrtc/...` and enforces the existing operator JWT + RBAC there.
If you only ever want P2P, you can comment out the `mediamtx:` service in your compose file — only Relayed mode depends on it.
MediaMTX automatically discovers addresses assigned directly to its network interfaces. This covers ordinary Linux host-network and many LAN deployments, but it cannot reliably infer a public address created by NAT, port forwarding, a cloud load balancer, or a reverse proxy. It also cannot discover which public DNS name you intend operators to use.
- **Linux with host networking and no public NAT:** normally nothing is required; MediaMTX sees the host's interfaces automatically.
- **Public server behind NAT, port forwarding, or a cloud public-IP mapping:** set the public IP or a DNS name that resolves to it.
- **Windows / macOS Docker Desktop:** set the host's reachable LAN/public address because automatic interface discovery sees the container network, not necessarily the Docker host.
- **Same-machine testing on Docker Desktop:** retain `127.0.0.1` in the list.
Ensure inbound UDP `8189` is forwarded to the MediaMTX host/container; TCP `8189` is the slower fallback. Recreate the sidecar after changing the setting:
```bash
# Linux
docker compose up -d --force-recreate mediamtx
# Windows / macOS Docker Desktop
docker compose -f docker-compose.windows.yml up -d --force-recreate mediamtx
```
To verify the selected route, open `chrome://webrtc-internals` during a relayed stream and confirm that the chosen remote candidate points to the expected server address and preferably uses UDP.
Automatic public-IP discovery is possible with STUN, but it is not a complete replacement for this setting: it can require random UDP ports, does not discover your intended domain, and can fail behind restrictive or symmetric NAT. For predictable production deployments, explicitly setting the reachable IP/domain and forwarding fixed UDP port `8189` is recommended.
On a Linux test host, `scripts/rd-network-impairment.sh` applies repeatable bandwidth, latency, jitter, and packet-loss profiles with `tc netem`. Give it the exact interface carrying the WebRTC traffic:
```bash
# Inspect interfaces first; do not guess the target.
ip link show
sudo bash scripts/rd-network-impairment.sh apply eth0 wan
Available profiles are `wan` (20 Mbps), `constrained` (8 Mbps), and `harsh` (3 Mbps), with progressively higher delay, jitter, and loss. The rule affects all egress traffic on the selected interface, not only Overlord. The script refuses to touch loopback unless `OVERLORD_NETEM_ALLOW_LOOPBACK=1` is explicitly set.
For each profile, start Remote Desktop with both **Profile: Auto (best)** and **Bitrate: Auto**, then watch the Stats HUD. Verify that bitrate falls before the profile steps down, decode/processing queues remain bounded, and the profile slowly recovers after clearing the impairment.
The compose file passes a minimal set of `MTX_*` environment variables to MediaMTX — enough for Overlord to work. To change anything else (codecs, paths, ICE servers, etc.), either:
- Add more `MTX_*` variables to the `mediamtx` service's `environment:` block (every option in MediaMTX's docs is supported as an env var), or
- Provide a full `mediamtx.yml` via a bind mount:
```yaml
mediamtx:
# ...existing config...
volumes:
- ./mediamtx.yml:/mediamtx.yml:ro
```
Make sure the file exists on the host *before* the container starts — Docker will otherwise auto-create an empty directory at that path and fail with "not a directory".
Overlord supports generic OIDC login for homelab identity providers such as Authentik, Authelia, Keycloak, Zitadel, and Dex. Local username/password login stays enabled as a fallback.
Configure your OIDC provider with this redirect URI:
New OIDC users are linked by the provider's `issuer + sub` identity. Email/username linking to an existing local account is disabled by default; enable `OVERLORD_OIDC_ALLOW_EMAIL_LINK=true` only if you trust your provider's verified email claims.
---
## Login Branding
Self-hosted and enterprise deployments can brand the login screen with environment variables:
Logo, hero, and support URLs accept absolute `http(s)` URLs or root-relative paths. If no logo URL is set, the UI uses `OVERLORD_LOGIN_ICON_CLASS`. The same options can also be managed from Settings → Branding.
`docker-compose.yml` ships with `build.cache_from` and `build.cache_to` pointing at `.docker-cache/buildx`. Local builds reuse it automatically — no extra setup.
When a macOS target is selected with CGO enabled, the builder asks for a user-provided macOS SDK archive. Package the complete `MacOSX*.sdk` directory as `.tar.xz`, `.tar.gz`, `.tgz`, or `.tar` and upload it from the Build page. The archive must contain the SDK's `System/Library/Frameworks` and `usr` directories.
The upload is limited to 1 GB, belongs to the authenticated user, can be used for only one build, and is deleted after the build. Unused uploads expire after one hour. The Docker image supplies Clang and LLD; Apple SDK files are never bundled or downloaded by Overlord.
If the dashboard, audit log, or IP bans show all agents as `172.x.x.x` (or some other proxy/bridge IP), something between the agent and Bun is rewriting the source IP. Two independent flags govern this:
- `OVERLORD_TLS_OFFLOAD` — TLS terminates at the proxy; Overlord runs plain HTTP internally.
- `OVERLORD_TRUST_PROXY` — honor `X-Forwarded-For` / `X-Real-IP` / `CF-Connecting-IP` so dashboard/audit/IP-bans see the real client. Auto-enabled when `TLS_OFFLOAD=true`.
| Domain, proxy does TLS (Render, nginx terminating TLS → http upstream) | `true` | auto (`true`) |
| Domain, proxy in front but Overlord still does TLS (Cloudflare orange-cloud Full Strict; nginx with `proxy_pass https://`) | `false` | `true` |
Only enable `OVERLORD_TRUST_PROXY` when a trusted reverse proxy is in front. If Overlord is directly exposed and you enable it, agents can spoof their source IP by sending their own `X-Forwarded-For` header, breaking IP bans and audit accuracy. The upstream proxy also has to be configured to inject the header (Cloudflare does by default; nginx/Caddy/Traefik need explicit directives).
If you're using `docker-compose.windows.yml` or `docker-compose.quickstart.yml` (Docker Desktop bridge networking) with **no** reverse proxy, Docker itself rewrites source IPs to the bridge gateway and there is no header to recover from — `TRUST_PROXY` cannot help. Either switch to Linux host networking or put a real reverse proxy in front.
