Add Vercel sandbox environment support

README.md

@@ -77,11 +77,11 @@ make quickstart
 </details>
 
 ## REPL Environments
-We support two types of REPL environments -- isolated, and non-isolated. Non-isolated environments (default) run code execution on the same machine as the RLM (e.g. through `exec`), which is pretty reasonable for some local low-risk tasks, like simple benchmarking, but can be problematic if the prompts or tool calls can interact with malicious users. Fully isolated environments use cloud-based sandboxes (e.g. Prime Sandboxes, [Modal Sandboxes](https://modal.com/docs/guide/sandboxes)) to run code generated by the RLM, ensuring complete isolation from the host process. Environments can be added, but we natively support the following: `local` (default), `docker`, `modal`, `prime`, `daytona`, `e2b`.
+We support two types of REPL environments -- isolated, and non-isolated. Non-isolated environments (default) run code execution on the same machine as the RLM (e.g. through `exec`), which is pretty reasonable for some local low-risk tasks, like simple benchmarking, but can be problematic if the prompts or tool calls can interact with malicious users. Fully isolated environments use cloud-based sandboxes (e.g. Prime Sandboxes, [Modal Sandboxes](https://modal.com/docs/guide/sandboxes), [Vercel Sandboxes](https://vercel.com/docs/vercel-sandbox)) to run code generated by the RLM, ensuring complete isolation from the host process. Environments can be added, but we natively support the following: `local` (default), `docker`, `modal`, `prime`, `daytona`, `e2b`, `vercel`.
 
 ```python
 rlm = RLM(
-    environment="...", # "local", "docker", "modal", "prime", "daytona", "e2b"
+    environment="...", # "local", "docker", "modal", "prime", "daytona", "e2b", "vercel"
     environment_kwargs={...},
 )
 ```
@@ -113,6 +113,21 @@ uv pip install -e ".[prime]"
 export PRIME_API_KEY=...
 ```
 
+#### Vercel Sandboxes
+To use [Vercel Sandboxes](https://vercel.com/docs/vercel-sandbox), install the SDK and authenticate with Vercel:
+```bash
+uv pip install -e ".[vercel]"
+vercel link
+vercel env pull
+```
+
+The Vercel SDK uses credentials stored in `.env.local` for local development. In Python examples, load it explicitly with `load_dotenv(".env.local")`.
+
+> [!NOTE]
+> `backend="vercel"` and `environment="vercel"` are different features:
+> - `backend="vercel"` uses the Vercel AI Gateway as an LM provider.
+> - `environment="vercel"` runs generated Python code inside Vercel Sandboxes.
+
 
 ### Model Providers
 We currently support most major clients (OpenAI, Anthropic), as well as the router platforms (OpenRouter, Portkey). For local models, we recommend using vLLM (which interfaces with the [OpenAI client](https://github.com/alexzhang13/rlm/blob/main/rlm/clients/openai.py)). To view or add support for more clients, start by looking at [`rlm/clients/`](https://github.com/alexzhang13/rlm/tree/main/rlm/clients).

examples/vercel_repl_example.py

@@ -0,0 +1,31 @@
+import os
+
+from dotenv import load_dotenv
+
+from rlm import RLM
+from rlm.logger import RLMLogger
+
+# Vercel Sandboxes local auth typically lives in .env.local after:
+#   vercel link
+#   vercel env pull
+load_dotenv(".env.local")
+
+logger = RLMLogger(log_dir="./logs")
+
+rlm = RLM(
+    backend="openai",
+    backend_kwargs={
+        "api_key": os.getenv("OPENAI_API_KEY"),
+        "model_name": "gpt-5-nano",
+    },
+    environment="vercel",
+    environment_kwargs={
+        "timeout": 300,
+    },
+    max_depth=1,
+    logger=logger,
+    verbose=True,
+)
+
+result = rlm.completion("Using your code, solve 2^(2^(2^(2))). Show your work in Python.")
+print(result.response)

pyproject.toml

@@ -39,6 +39,7 @@ modal = ["modal>=0.73.0", "dill>=0.3.7"]
 e2b = ["e2b-code-interpreter>=0.0.11", "dill>=0.3.7"]
 daytona = ["daytona>=0.128.1", "dill>=0.3.7"]
 prime = ["prime-sandboxes>=0.2.0", "dill>=0.3.7"]
+vercel = ["vercel>=0.4.0"]
 
 [build-system]
 requires = ["setuptools>=61.0"]

rlm/core/types.py

@@ -12,7 +12,7 @@
     "azure_openai",
     "gemini",
 ]
-EnvironmentType = Literal["local", "docker", "modal", "prime", "daytona", "e2b"]
+EnvironmentType = Literal["local", "docker", "modal", "prime", "daytona", "e2b", "vercel"]
 
 
 def _serialize_value(value: Any) -> Any:

rlm/environments/__init__.py

@@ -31,12 +31,12 @@
 
 
 def get_environment(
-    environment: Literal["local", "modal", "docker", "daytona", "prime", "e2b"],
+    environment: Literal["local", "modal", "docker", "daytona", "prime", "e2b", "vercel"],
     environment_kwargs: dict[str, Any],
 ) -> BaseEnv:
     """
     Routes a specific environment and the args (as a dict) to the appropriate environment if supported.
-    Currently supported environments: ['local', 'modal', 'docker', 'daytona', 'prime', 'e2b']
+    Currently supported environments: ['local', 'modal', 'docker', 'daytona', 'prime', 'e2b', 'vercel']
     """
     if environment == "local":
         return LocalREPL(**environment_kwargs)
@@ -60,7 +60,11 @@ def get_environment(
         from rlm.environments.e2b_repl import E2BREPL
 
         return E2BREPL(**environment_kwargs)
+    elif environment == "vercel":
+        from rlm.environments.vercel_repl import VercelREPL
+
+        return VercelREPL(**environment_kwargs)
     else:
         raise ValueError(
-            f"Unknown environment: {environment}. Supported: ['local', 'modal', 'docker', 'daytona', 'prime', 'e2b']"
+            f"Unknown environment: {environment}. Supported: ['local', 'modal', 'docker', 'daytona', 'prime', 'e2b', 'vercel']"
         )