diff --git a/README.md b/README.md
index af457d1..d04aa73 100644
--- a/README.md
+++ b/README.md
@@ -125,10 +125,11 @@ routstrd start --host 0.0.0.0
 
 Only expose the daemon behind appropriate network controls.
 
-With specific provider:
+Pin all requests to one provider (no cross-provider failover):
 ```sh
 routstrd start --provider https://your-provider.com
 ```
+See [Provider pinning](#provider-pinning) for the request-level header/query.
 
 ### CLI Commands
 
@@ -233,6 +234,23 @@ Any unmatched `POST` path is proxied to the selected provider with the incoming
 path preserved, so `POST /v1/messages` (Anthropic Messages API) and
 `POST /v1/responses` (OpenAI Responses API) work in their own formats too.
 
+#### Provider pinning
+
+Pin a request to a single provider with either the `x-routstr-provider` header
+or the `?provider=` query parameter (the query takes precedence):
+
+```sh
+curl -H 'x-routstr-provider: https://your-provider.com' ... \
+  http://127.0.0.1:8008/v1/chat/completions
+```
+
+A pin is strict: the request is sent only to that provider. On an upstream
+error — including a 400/422 the node itself rejects — the request is **not**
+failed over to another node; the error is returned to the client. Retries
+against the same provider (top-up, mint fallback) still happen. The config
+`provider` and `routstrd start --provider` set the same strict pin as a
+default for every request.
+
 Request body:
 ```json
 {
@@ -292,6 +310,10 @@ choose and pin an advertised model path for `deepseek-v4.1-flash` requests.
 Explicit `x-routstr-model-path` request headers work independently of this
 setting and take precedence. Restart the daemon after changing `autoModelPath`.
 
+`provider` pins every request to one node (same strict behavior as the
+`x-routstr-provider` header / `?provider=` query). Leave it `null` to let the
+router pick the cheapest eligible provider and fail over between nodes.
+
 ### Environment Variables
 
 - `ROUTSTRD_DIR` - Config directory (default: `~/.routstrd`)
diff --git a/SKILL.md b/SKILL.md
index 84a135b..69f5175 100644
--- a/SKILL.md
+++ b/SKILL.md
@@ -49,7 +49,7 @@ Start the background daemon process.
 |--------|-------------|
 | `--port <port>` | Port to listen on (default: 8008) |
 | `--host <host>` | Bind address (default: 127.0.0.1) |
-| `-p, --provider <provider>` | Default provider to use |
+| `-p, --provider <provider>` | Pin all requests to this provider (no cross-provider failover) |
 
 ### `routstrd daemon`
 
@@ -479,6 +479,20 @@ The incoming request path is forwarded to the provider, so the Anthropic
 Messages API (`POST /v1/messages`) and the OpenAI Responses API
 (`POST /v1/responses`) are proxied in their own formats as well.
 
+#### Provider pinning
+
+A provider can be pinned with any of:
+
+- request header `x-routstr-provider: <provider-url>`
+- query parameter `?provider=<provider-url>` (takes precedence over the header)
+- config `provider` / `routstrd start --provider <provider-url>`
+
+All three are **strict pins**: the request is routed only to that provider. If
+it answers with an error (including a 400/422 upstream rejection), the request
+is **not** failed over to another node — the error is returned to the client.
+Retries against the same provider (top-up, mint fallback) still happen. Without
+a pin, the router falls over to the next-cheapest eligible provider as usual.
+
 ## Configuration
 
 Config file: `~/.routstrd/config.json`
@@ -487,7 +501,7 @@ Config file: `~/.routstrd/config.json`
 |-------|------|---------|-------------|
 | `port` | number | 8008 | Daemon HTTP port |
 | `host` | string | `"127.0.0.1"` | Bind address |
-| `provider` | string\|null | null | Default provider URL |
+| `provider` | string\|null | null | Pinned provider URL; all requests go only to this node (no cross-provider failover) |
 | `mode` | string | `"apikeys"` | Client mode (`apikeys` or `xcashu`) |
 | `maxTokens` | number | 64000 | Completion budget applied when a client sets no output-token limit |
 | `daemonUrl` | string | — | Remote daemon URL (set by `routstrd remote`) |