diff --git a/README.md b/README.md
index 8a6c1ccd..9b8934a6 100644
--- a/README.md
+++ b/README.md
@@ -28,6 +28,7 @@ Create a `.env` file in the project root (or export these in your shell):
 ```bash
 ANTHROPIC_API_KEY=<your-anthropic-api-key> # if using anthropic models
 OPENAI_API_KEY=<your-openai-api-key> # if using openai models
+OPENROUTER_API_KEY=<your-openrouter-api-key> # if using openrouter models
 HF_TOKEN=<your-hugging-face-token>
 GITHUB_TOKEN=<github-personal-access-token> 
 ```
@@ -52,10 +53,15 @@ ml-intern "fine-tune llama on my dataset"
 ```bash
 ml-intern --model anthropic/claude-opus-4-6 "your prompt"
 ml-intern --model openai/gpt-5.5 "your prompt"
+ml-intern --model openrouter/anthropic/claude-sonnet-4-5 "your prompt"
 ml-intern --max-iterations 100 "your prompt"
 ml-intern --no-stream "your prompt"
 ```
 
+OpenRouter models use the `openrouter/<provider>/<model>` naming convention and
+the `OPENROUTER_API_KEY` environment variable. Optional attribution headers can
+be set with `OR_SITE_URL` and `OR_APP_NAME`.
+
 ## Supported Gateways
 
 ML Intern currently supports one-way notification gateways from CLI sessions.
diff --git a/agent/core/llm_params.py b/agent/core/llm_params.py
index 880886b3..b1b17126 100644
--- a/agent/core/llm_params.py
+++ b/agent/core/llm_params.py
@@ -5,6 +5,8 @@
 creating circular imports.
 """
 
+import os
+
 from agent.core.hf_tokens import get_hf_bill_to, resolve_hf_router_token
 
 
@@ -72,12 +74,14 @@ def _widened(model: str) -> bool:
 # Effort levels accepted on the wire.
 #   Anthropic (4.6+):  low | medium | high | xhigh | max   (output_config.effort)
 #   OpenAI direct:     minimal | low | medium | high | xhigh (reasoning_effort top-level)
+#   OpenRouter:        minimal | low | medium | high | xhigh (reasoning_effort top-level)
 #   HF router:         low | medium | high                 (extra_body.reasoning_effort)
 #
 # We validate *shape* here and let the probe cascade walk down on rejection;
 # we deliberately do NOT maintain a per-model capability table.
 _ANTHROPIC_EFFORTS = {"low", "medium", "high", "xhigh", "max"}
 _OPENAI_EFFORTS = {"minimal", "low", "medium", "high", "xhigh"}
+_OPENROUTER_EFFORTS = {"minimal", "low", "medium", "high", "xhigh"}
 _HF_EFFORTS = {"low", "medium", "high"}
 
 
@@ -114,6 +118,10 @@ def _resolve_llm_params(
     • ``openai/<model>`` — ``reasoning_effort`` forwarded as a top-level
       kwarg (GPT-5 / o-series). LiteLLM uses the user's ``OPENAI_API_KEY``.
 
+    • ``openrouter/<provider>/<model>`` — routed through LiteLLM's OpenRouter
+      provider. LiteLLM uses ``OPENROUTER_API_KEY``; optional attribution
+      headers can be set with ``OR_SITE_URL`` and ``OR_APP_NAME``.
+
     • Anything else is treated as a HuggingFace router id. We hit the
       auto-routing OpenAI-compatible endpoint at
       ``https://router.huggingface.co/v1``. The id can be bare or carry an
@@ -180,6 +188,31 @@ def _resolve_llm_params(
                 params["reasoning_effort"] = reasoning_effort
         return params
 
+    if model_name.startswith("openrouter/"):
+        params = {"model": model_name}
+        if api_key := os.getenv("OPENROUTER_API_KEY", "").strip():
+            params["api_key"] = api_key
+        if api_base := os.getenv("OPENROUTER_API_BASE", "").strip():
+            params["api_base"] = api_base
+
+        extra_headers = {}
+        if site_url := os.getenv("OR_SITE_URL", "").strip():
+            extra_headers["HTTP-Referer"] = site_url
+        if app_name := os.getenv("OR_APP_NAME", "").strip():
+            extra_headers["X-Title"] = app_name
+        if extra_headers:
+            params["extra_headers"] = extra_headers
+
+        if reasoning_effort:
+            if reasoning_effort not in _OPENROUTER_EFFORTS:
+                if strict:
+                    raise UnsupportedEffortError(
+                        f"OpenRouter doesn't accept effort={reasoning_effort!r}"
+                    )
+            else:
+                params["reasoning_effort"] = reasoning_effort
+        return params
+
     hf_model = model_name.removeprefix("huggingface/")
     api_key = _resolve_hf_router_token(session_hf_token)
     params = {
diff --git a/tests/unit/test_llm_params.py b/tests/unit/test_llm_params.py
index 5234461a..262c7528 100644
--- a/tests/unit/test_llm_params.py
+++ b/tests/unit/test_llm_params.py
@@ -30,6 +30,47 @@ def test_openai_max_effort_is_still_rejected():
         raise AssertionError("Expected UnsupportedEffortError for max effort")
 
 
+def test_openrouter_params_use_openrouter_provider(monkeypatch):
+    monkeypatch.setenv("OPENROUTER_API_KEY", " openrouter-token ")
+    monkeypatch.setenv("OR_SITE_URL", " https://example.com/ml-intern ")
+    monkeypatch.setenv("OR_APP_NAME", " ML Intern ")
+
+    params = _resolve_llm_params(
+        "openrouter/anthropic/claude-sonnet-4-5",
+        reasoning_effort="high",
+        strict=True,
+    )
+
+    assert params["model"] == "openrouter/anthropic/claude-sonnet-4-5"
+    assert params["api_key"] == "openrouter-token"
+    assert params["reasoning_effort"] == "high"
+    assert params["extra_headers"] == {
+        "HTTP-Referer": "https://example.com/ml-intern",
+        "X-Title": "ML Intern",
+    }
+
+
+def test_openrouter_api_base_override(monkeypatch):
+    monkeypatch.setenv("OPENROUTER_API_BASE", " https://openrouter.example/v1 ")
+
+    params = _resolve_llm_params("openrouter/openai/gpt-5.5")
+
+    assert params["api_base"] == "https://openrouter.example/v1"
+
+
+def test_openrouter_rejects_max_effort_in_strict_mode():
+    try:
+        _resolve_llm_params(
+            "openrouter/openai/gpt-5.5",
+            reasoning_effort="max",
+            strict=True,
+        )
+    except UnsupportedEffortError as exc:
+        assert "OpenRouter doesn't accept effort='max'" in str(exc)
+    else:
+        raise AssertionError("Expected UnsupportedEffortError for max effort")
+
+
 def test_hf_router_token_prefers_inference_token(monkeypatch):
     monkeypatch.setenv("INFERENCE_TOKEN", " inference-token ")
     monkeypatch.setenv("HF_TOKEN", "hf-token")