Create app.py

app.py

@@ -0,0 +1,140 @@
+import os
+from dotenv import load_dotenv
+from e2b import Sandbox
+from huggingface_hub import InferenceClient # Import the InferenceClient
+import gradio as gr
+
+# Load environment variables from .env file
+load_dotenv()
+
+# --- E2B Setup ---
+e2b_api_key = os.getenv("E2B_API_KEY")
+if not e2b_api_key:
+    print("WARNING: E2B_API_KEY not found. Cannot run locally without it.")
+
+# --- LLM (Hugging Face Inference API) Setup ---
+hf_token = os.getenv("HF_TOKEN") # Get Hugging Face token
+if not hf_token:
+    print("WARNING: HF_TOKEN not found. Cannot run locally without it. Inference might be limited.")
+
+# Choose the model you want to use from Hugging Face
+# Make sure it's a text generation or chat model.
+# Examples: "mistralai/Mistral-7B-Instruct-v0.2", "meta-llama/Llama-2-7b-chat-hf", "Qwen/Qwen1.5-7B-Chat"
+HF_MODEL_ID = "mistralai/Mistral-7B-Instruct-v0.2" # <-- **Choose your model here**
+
+# Initialize the Inference Client
+try:
+    llm_client = InferenceClient(model=HF_MODEL_ID, token=hf_token)
+    llm_client_available = True
+except Exception as e:
+    print(f"Could not initialize Hugging Face Inference Client: {e}")
+    llm_client_available = False
+    llm_client = None # Ensure client is None if initialization fails
+
+
+def run_agent_task(user_input: str):
+    """
+    Processes user input, interacts with Hugging Face Inference API and E2B, and returns results.
+    """
+    output = ""
+
+    if not e2b_api_key:
+        return "Error: E2B API key not configured."
+    if not llm_client_available or llm_client is None:
+         return "Error: Hugging Face Inference Client not initialized. Check HF_TOKEN or model ID."
+
+    # Start an E2B sandbox session
+    try:
+        with Sandbox(api_key=e2b_api_key, template="base") as sandbox:
+            output += "E2B Sandbox started successfully.\n"
+
+            # 1. Formulate a prompt for the LLM
+            # Use the same prompt structure, adapting slightly if needed for chat models
+            prompt_content = f"""
+            You are a computer agent connected to a sandboxed environment.
+            The user wants you to perform the following task: {user_input}
+
+            Based on the task, decide what command(s) to run in the bash terminal within the sandbox.
+            Output only the bash command(s), nothing else. If no command is needed, output 'NO_COMMAND'.
+            For example:
+            User: List files in the current directory
+            Agent: ls -l
+
+            User: Calculate 2+2
+            Agent: echo $((2+2))
+
+            User: Greet me
+            Agent: NO_COMMAND
+
+            Now, based on the user task: {user_input}
+            Output the bash command(s) or 'NO_COMMAND':
+            """
+
+            # 2. Call the Hugging Face Inference API
+            try:
+                # Use the .chat method which is suitable for instruction-following models
+                # It takes messages in the OpenAI format
+                response = llm_client.chat(
+                    messages=[
+                        {"role": "system", "content": "You are a helpful assistant that outputs bash commands or NO_COMMAND."},
+                        {"role": "user", "content": prompt_content}
+                    ],
+                    max_tokens=100, # Adjust as needed
+                    temperature=0.1, # Lower temperature often helps with predictable output like commands
+                    # Add other parameters as needed (e.g., top_p)
+                )
+
+                # Extract the content from the response
+                # The response structure depends on the method and model, .chat returns a ChatCompletion object
+                command_to_run = response.choices[0].message.content.strip()
+
+                output += f"LLM ({HF_MODEL_ID}) decided to run: `{command_to_run}`\n"
+
+            except Exception as e:
+                output += f"An exception occurred calling Hugging Face Inference API: {e}\n"
+                command_to_run = "NO_COMMAND" # Prevent execution on LLM error
+
+            # 3. Execute the command in the E2B sandbox (if not NO_COMMAND)
+            # This part remains the same as it uses the E2B SDK
+            if command_to_run and command_to_run != "NO_COMMAND":
+                try:
+                    proc = sandbox.process.start(command_to_run)
+                    process_output = proc.wait()
+
+                    if process_output.stdout:
+                        output += "--- Command Output (stdout) ---\n"
+                        output += process_output.stdout + "\n"
+                    if process_output.stderr:
+                        output += "--- Command Output (stderr) ---\n"
+                        output += process_output.stderr + "\n"
+
+                    output += f"Command exited with code: {process_output.exit_code}\n"
+
+                except Exception as e:
+                    output += f"Error executing command in sandbox: {e}\n"
+            elif command_to_run == "NO_COMMAND":
+                 output += "LLM decided no command was necessary.\n"
+            else:
+                 output += "LLM returned an empty command.\n"
+
+            output += "E2B Sandbox session ended.\n"
+
+    except Exception as e:
+        output += f"An error occurred with the E2B sandbox: {e}\n"
+        output += "Please check your E2B API key and try again.\n"
+
+    return output
+
+
+# Define the Gradio interface - This part is unchanged
+interface = gr.Interface(
+    fn=run_agent_task,
+    inputs=gr.Textbox(lines=2, placeholder="Enter your task for the agent here..."),
+    outputs=gr.Textbox(lines=10, label="Agent Output", interactive=False),
+    title=f"E2B Computer Agent Demo (using {HF_MODEL_ID})", # Update title
+    description="Enter a task for the agent to perform in a sandboxed environment using E2B and a Hugging Face model via Inference API.",
+)
+
+# This is the line Hugging Face Spaces will look for
+if __name__ == "__main__":
+    interface.launch()