diff --git a/crates/llm/src/bin/ullm.rs b/crates/llm/src/bin/ullm.rs
index 3b68ba2a4..3c9701446 100644
--- a/crates/llm/src/bin/ullm.rs
+++ b/crates/llm/src/bin/ullm.rs
@@ -9,6 +9,10 @@ use llm::generate::{self, GenerateParams};
 #[derive(Parser)]
 #[command(name = "ullm")]
 struct Cli {
+    /// Skip loading .env file
+    #[arg(long, global = true)]
+    no_dotenv: bool,
+
     #[command(subcommand)]
     command: Command,
 }
@@ -217,8 +221,10 @@ async fn run_prompt(args: PromptArgs) -> Result<()> {
 
 #[tokio::main]
 async fn main() -> Result<()> {
-    dotenvy::dotenv().ok();
     let cli = Cli::parse();
+    if !cli.no_dotenv {
+        dotenvy::dotenv().ok();
+    }
 
     match cli.command {
         Command::Prompt {
@@ -367,10 +373,8 @@ mod tests {
     fn prompt_reads_from_stdin() {
         // This test verifies stdin is read, but will fail at the API call stage
         // since no API key is set. The error should NOT be "no prompt provided".
-        // env_clear() prevents .env loading from triggering real API calls.
         let result = ullm()
-            .env_clear()
-            .args(["prompt", "--no-stream", "-m", "test-model"])
+            .args(["--no-dotenv", "prompt", "--no-stream", "-m", "test-model"])
             .write_stdin("hello from stdin")
             .assert()
             .failure();
@@ -383,10 +387,8 @@ mod tests {
     #[test]
     fn prompt_concatenates_stdin_and_arg() {
         // Same as above — verifies it doesn't error on "no prompt"
-        // env_clear() prevents .env loading from triggering real API calls.
         let result = ullm()
-            .env_clear()
-            .args(["prompt", "--no-stream", "-m", "test-model", "summarize this"])
+            .args(["--no-dotenv", "prompt", "--no-stream", "-m", "test-model", "summarize this"])
             .write_stdin("some input text")
             .assert()
             .failure();