The unshackle REST API protects every endpoint (with one exception) behind a single API key. Clients present that API key in a request header, and the server examines it against the API keys in your configuration. This page explains how that check works, how to configure API keys, and how clients send them. It also explains the extra controls on top of the API key: per-key service allowlists, per-key CDM device access, and remote session IP binding.
You start the API with the [`unshackle serve`](../../guide/cli-reference.md) command. Authentication is on by default: if you start the server without an API key configured, it refuses to launch.
The header, the config, and the behaviour you see as a client are the same in both. The differences are internal, and the developer notes below give them.
This value becomes the one and only API key that the server accepts. The server allows requests that present it in the `X-Secret-Key` header. The server rejects everything else.
For finer control you can define named users, each with their own API key. A user entry can also grant CDM device access and restrict which services that API key may touch:
The config key of each entry under `users` (for example `a-secret-key-for-alice`) *is* the secret that client sends in `X-Secret-Key`. The `username` is a human-readable label used in logs and in job/history metadata.
On the integrated server, the server automatically adds the configured `api_secret` as an extra accepted API key alongside everyone in `users`. It gives that API key access to all loaded Widevine and PlayReady devices, and labels it `api_user`. So `api_secret` behaves as an admin-level API key in addition to any named users you define.
If a user entry omits `playready_devices`, the server gives that API key **no** PlayReady access, and logs a warning at startup. You must name the PlayReady devices explicitly for any API key that needs them. Widevine and PlayReady device files themselves are auto-loaded from your WVDs and PRDs directories.
There is no separate tier, capability flag, or permission toggle for whether the server will hand back content keys (`KID:KEY`) for a remote session download. The presence of devices on the calling API key decides it. With **empty** `devices` and `playready_devices`, the server can only proxy CDM challenges, so the client must operate its own CDM. Once those lists are **populated**, the client may request `mode:server_cdm` and receive content keys back. Do not look for a tier setting. The devices you configure are what enables server-side decryption.
The `--no-key` flag turns authentication off entirely. The server allows every request, ignores `X-Secret-Key`, and needs no API key in your configuration.
With `--no-key`, any client that can reach the port can show your services, start downloads, and use your CDM devices. Never put a `--no-key` server on an untrusted network.
The value must match one of the configured API keys exactly (`api_secret`, or one of the keys under `users`). There is no login step, no HTTP session cookie, and no token exchange for API access. The client sends the same static API key on each request.
The job event stream, `GET /api/download/jobs/{job_id}/events`, also accepts the API key in a `secret_key` query parameter. This is for browsers: the `EventSource` API cannot set request headers. The exception applies to that one route only, and it works in both server modes.
The header always wins. If the request has an `X-Secret-Key` header, the server examines that header and ignores the query parameter. The server rejects a wrong header even when the query parameter holds a valid API key.
An API key in a URL is easier to leak than an API key in a header, because proxies and browser history can record it. Use the header wherever your client can set one.
unshackle's own [remote download](../../guide/downloading.md) client (`RemoteService`) sends `X-Secret-Key` automatically, reading the API key from your `remote_services` config. It also sends a `User-Agent` of `unshackle/<version>`. You only deal with the header directly when writing your own client.
One endpoint is exempt from authentication: `GET /api/health`. It never requires an API key, so you can use it for liveness and version probes without embedding a secret in your monitoring:
!!! warning "Auth errors use a different JSON shape than other errors (developer note)"
The rest of the API reports errors with a structured body containing `error_code`, `message`, `timestamp`, and other fields (see [Errors](errors.md)). The authentication middleware is different: it returns `{"status": 401, "message": "..."}`, where `status` is an **integer**, and there is no `error_code` or `timestamp`.
Endpoints that name a service (listing services, searching, listing titles/tracks, downloading, profiles, history, and remote sessions) all filter against this allowlist. The server rejects a request for a service outside the API key's allowlist as an invalid service. The server matches service tags after normalisation, so casing and aliases give the same result.
The server config gates two download-time capabilities. Unless you enable them, the server refuses them with `403 Forbidden`, even for an authenticated API key:
| `serve.cdm_overrides` | off | Whether a request may pick a specific CDM device via the `cdm` field. Set to `true` to allow any, or to a list/set of device names to allow only those. |
| `serve.allow_job_credentials` | off | Whether a request may pass its own `credential` / `credentials` login details for a download job. |
```yaml title="Opting in to per-request CDM and credentials"
serve:
api_secret: "admin-key"
cdm_overrides: ["my_widevine_device"] # or true to allow any configured device
The two gates are fully independent, and you can combine them. For a locked-down or public-facing deployment, leave **both** unset. They then default off and reject with `403`, so client requests can never steer the server's CDM or feed it their own logins. Enable them only for a trusted single-client deployment, where the caller and the operator are effectively the same party.
When you enable `allow_job_credentials` and a download job supplies its own `credential` / `credentials`, the server gives each distinct credential its **own** isolated token cache. The server never shares or merges client-supplied logins with its configured-credential caches. They cannot cross-contaminate or read the server's own token state, and they do not clobber each other.
The remote-download session endpoints (`/api/session/*`) add an extra check on top of the API key. When the server creates a remote session, it records the client's IP address. Any later request for that remote session must come from the **same** IP, or the server rejects it with `403 Forbidden`. Thus the server binds a remote session both to the API key that created it and to the address it came from.
The API server speaks plain HTTP. Because the client sends your API key on every request, do not use it over an unencrypted connection on an untrusted network. To serve over HTTPS, operate a reverse proxy in front of unshackle. The `serve` command can launch Caddy for you:
With `--caddy`, unshackle also starts `caddy run` using the `Caddyfile` located next to your unshackle config, letting Caddy end the TLS connection and reverse-proxy to the API. Configuring Caddy itself (certificates, domains, upstream) is done in that `Caddyfile`.
- **API-only server** (`--api-only`, or `--remote-only` which forces API-only). An `api_key_authentication` middleware runs on every request. It lets `GET /api/health` through unconditionally, then reads `X-Secret-Key`. An empty header returns the *"Secret Key is Empty."* body. A header that is not a configured user returns *"Secret Key is Invalid."*. In this mode the accepted-keys map holds your `api_secret` (labelled `api_user`, with no devices) plus every key under `serve.users`. A per-user key therefore authenticates here exactly as it does on the integrated server, and that user's `services` allowlist applies as usual. With `--no-key`, this middleware is not installed and the users map is empty.
- **Integrated server** (default). A `serve_authentication` middleware wraps pywidevine's authentication (and pyplayready's for `/playready*` paths), and the REST routes live in the same application, guarded by the same check. Here the accepted-keys map is your full `serve` config: `api_secret` (auto-added with access to all loaded devices) plus every entry under `users`, each with its declared `devices` / `playready_devices`.
Because the API-only middleware guards *all* non-health paths, the Swagger UI at `/api/docs/` is also behind `X-Secret-Key` there. In `--remote-only` mode the Swagger UI is not mounted at all.
Both middlewares read the caller's API key through the shared `request_secret_key()` helper in `unshackle/core/api/handlers.py`. That helper returns the `X-Secret-Key` header, and falls back to the `secret_key` query parameter only when the matched route is the job event stream. The integrated server needs one extra branch for this. pywidevine's `authentication` is header-only. So `serve_authentication` compares the API key from the query parameter itself when the header is absent, and delegates to pywidevine in every other case. That is why both modes reject a present-but-wrong header.
The REST API authenticates with `X-Secret-Key`. The middleware advertises `Authorization` because the co-hosted pywidevine / pyplayready CDM endpoints use it. `PUT` appears in the allowed-methods list, but no REST route uses it. Only `GET`, `POST`, and `DELETE` have routes.
The map of accepted API keys lives at `request.app["config"]["users"]`. For DRM operations, an API key's permitted Widevine devices come from its `devices` list, and PlayReady devices from `playready_devices`. The server uses the first configured device when a request does not (or may not) override it. A server-wide `config.cdm[<service>]` mapping, if present, takes precedence over the API key's device list when the server chooses which device to use.
