diff --git a/poc-memory/src/agents/api.rs b/poc-memory/src/agents/api.rs
index acdffe1..4a25b10 100644
--- a/poc-memory/src/agents/api.rs
+++ b/poc-memory/src/agents/api.rs
@@ -12,6 +12,20 @@ use poc_agent::types::*;
 use poc_agent::tools::{self, ProcessTracker};
 use poc_agent::ui_channel::StreamTarget;
 
+use std::sync::OnceLock;
+
+static API_CLIENT: OnceLock<ApiClient> = OnceLock::new();
+
+fn get_client() -> Result<&'static ApiClient, String> {
+    Ok(API_CLIENT.get_or_init(|| {
+        let config = crate::config::get();
+        let base_url = config.api_base_url.as_deref().unwrap_or("");
+        let api_key = config.api_key.as_deref().unwrap_or("");
+        let model = config.api_model.as_deref().unwrap_or("qwen-2.5-27b");
+        ApiClient::new(base_url, api_key, model)
+    }))
+}
+
 /// Run an agent prompt through the direct API with tool support.
 /// Returns the final text response after all tool calls are resolved.
 pub async fn call_api_with_tools(
@@ -19,14 +33,7 @@ pub async fn call_api_with_tools(
     prompt: &str,
     log: &dyn Fn(&str),
 ) -> Result<String, String> {
-    let config = crate::config::get();
-
-    let base_url = config.api_base_url.as_deref()
-        .ok_or("api_base_url not configured")?;
-    let api_key = config.api_key.as_deref().unwrap_or("");
-    let model = config.api_model.as_deref().unwrap_or("qwen-2.5-27b");
-
-    let client = ApiClient::new(base_url, api_key, model);
+    let client = get_client()?;
 
     // Set up a minimal UI channel (we just collect messages, no TUI)
     let (ui_tx, _ui_rx) = poc_agent::ui_channel::channel();
@@ -104,18 +111,15 @@ pub async fn call_api_with_tools(
 pub fn call_api_with_tools_sync(
     agent: &str,
     prompt: &str,
-    log: &dyn Fn(&str),
+    log: &(dyn Fn(&str) + Sync),
 ) -> Result<String, String> {
-    // Run on a new thread to avoid conflicts with any existing runtime
-    let agent = agent.to_string();
-    let prompt = prompt.to_string();
     std::thread::scope(|s| {
         s.spawn(|| {
             let rt = tokio::runtime::Builder::new_current_thread()
                 .enable_all()
                 .build()
                 .map_err(|e| format!("tokio runtime: {}", e))?;
-            rt.block_on(call_api_with_tools(&agent, &prompt, &|msg| eprintln!("[api] {}", msg)))
+            rt.block_on(call_api_with_tools(agent, prompt, log))
         }).join().unwrap()
     })
 }