diff --git a/crates/skald-core/src/llm/providers/mod.rs b/crates/skald-core/src/llm/providers/mod.rs
index 0be1b2b..d429ec8 100644
--- a/crates/skald-core/src/llm/providers/mod.rs
+++ b/crates/skald-core/src/llm/providers/mod.rs
@@ -48,13 +48,22 @@ pub(crate) fn extra_with_reasoning(
 /// the `data` envelope so each provider can map and enrich them with its own
 /// heuristics. `api_key` is sent as a bearer token when present (local
 /// providers pass `None`); `who` is the display name used in error messages.
+///
+/// `query` appends a raw query string (no leading `?`). Plain OpenAI has no
+/// filters, but a gateway hosting several service kinds needs one to say which
+/// catalogue it wants — OpenRouter's STT models are absent from the unfiltered
+/// listing and only appear under `output_modalities=transcription`.
 pub(crate) async fn fetch_openai_models(
     http:     &reqwest::Client,
     base_url: &str,
     api_key:  Option<&str>,
+    query:    Option<&str>,
     who:      &str,
 ) -> Result<Vec<serde_json::Value>> {
-    let url = format!("{}/models", base_url.trim_end_matches('/'));
+    let url = match query {
+        Some(q) => format!("{}/models?{q}", base_url.trim_end_matches('/')),
+        None    => format!("{}/models", base_url.trim_end_matches('/')),
+    };
     let mut req = http.get(&url);
     if let Some(key) = api_key {
         req = req.bearer_auth(key);
diff --git a/crates/skald-core/src/llm/providers/openrouter.rs b/crates/skald-core/src/llm/providers/openrouter.rs
index a3aae5e..6f1eb93 100644
--- a/crates/skald-core/src/llm/providers/openrouter.rs
+++ b/crates/skald-core/src/llm/providers/openrouter.rs
@@ -6,7 +6,7 @@ use crate::image_generate::ImageGenerateModelRecord;
 use crate::image_generate::openrouter_image::OpenRouterImageGenerator;
 use crate::llm::{LlmModelRecord, LlmProviderRecord};
 use crate::llm::providers::{RemoteLlmModelInfo, build_openai_llm, fetch_openai_models};
-use crate::transcribe::TranscribeModelRecord;
+use crate::transcribe::{RemoteTranscribeModelInfo, TranscribeModelRecord};
 use crate::transcribe::openai_audio::OpenAiAudioTranscriber;
 use crate::tts::TtsModelRecord;
 use crate::tts::openai_tts::OpenAiTtsSynthesiser;
@@ -47,7 +47,7 @@ impl OpenRouterProvider {
     }
 
     async fn fetch_catalog(&self, api_key: &str) -> Result<Vec<RemoteLlmModelInfo>> {
-        let raw = fetch_openai_models(&self.http, "https://openrouter.ai/api/v1", Some(api_key), "OpenRouter").await?;
+        let raw = fetch_openai_models(&self.http, "https://openrouter.ai/api/v1", Some(api_key), None, "OpenRouter").await?;
 
         let models = raw
             .iter()
@@ -97,6 +97,35 @@ impl OpenRouterProvider {
 
         Ok(models)
     }
+
+    /// OpenRouter's STT catalogue is the same `/models` envelope, filtered by
+    /// output modality. The filter is not an optimisation: these models carry
+    /// `architecture.modality = "audio->transcription"` and are **absent** from
+    /// the unfiltered listing, so nothing but this query surfaces them.
+    ///
+    /// The feed says nothing about which languages each model covers (every
+    /// entry here is multilingual or auto-detecting anyway), hence the empty
+    /// `languages` — the per-model hint stays the user's to set.
+    async fn fetch_transcribe_catalog(&self, api_key: &str) -> Result<Vec<RemoteTranscribeModelInfo>> {
+        let raw = fetch_openai_models(
+            &self.http, "https://openrouter.ai/api/v1", Some(api_key),
+            Some("output_modalities=transcription"), "OpenRouter",
+        ).await?;
+
+        Ok(raw
+            .iter()
+            .filter_map(|m| {
+                let id   = m["id"].as_str()?.to_string();
+                let name = m["name"].as_str().unwrap_or(&id).to_string();
+                Some(RemoteTranscribeModelInfo {
+                    id,
+                    name,
+                    description: m["description"].as_str().map(String::from),
+                    languages:   Vec::new(),
+                })
+            })
+            .collect())
+    }
 }
 
 #[async_trait::async_trait]
@@ -113,6 +142,12 @@ impl ApiProvider for OpenRouterProvider {
         Ok(Some(self.fetch_catalog(api_key).await?))
     }
 
+    async fn list_transcribe_models(&self, record: &LlmProviderRecord) -> Result<Option<Vec<RemoteTranscribeModelInfo>>> {
+        let api_key = record.api_key.as_deref()
+            .ok_or_else(|| anyhow!("provider '{}': api_key required for openrouter model listing", record.name))?;
+        Ok(Some(self.fetch_transcribe_catalog(api_key).await?))
+    }
+
     fn reasoning_mode(&self, _model_id: &str, capabilities: &[String]) -> Option<ReasoningMode> {
         // Fallback for stored/manually-added models with no catalog descriptor.
         // The precise per-model set comes from `parse_reasoning` in the catalog;
diff --git a/crates/skald-core/src/llm/providers/requesty.rs b/crates/skald-core/src/llm/providers/requesty.rs
index 601d06f..cdb6e7f 100644
--- a/crates/skald-core/src/llm/providers/requesty.rs
+++ b/crates/skald-core/src/llm/providers/requesty.rs
@@ -28,7 +28,7 @@ impl RequestyProvider {
     }
 
     async fn fetch_catalog(&self, api_key: &str) -> Result<Vec<RemoteLlmModelInfo>> {
-        let raw = fetch_openai_models(self.http(), BASE_URL, Some(api_key), "Requesty").await?;
+        let raw = fetch_openai_models(self.http(), BASE_URL, Some(api_key), None, "Requesty").await?;
         Ok(raw.iter().filter_map(map_model).collect())
     }
 }