diff --git a/internal/mcp/init_retry_test.go b/internal/mcp/init_retry_test.go
new file mode 100644
index 00000000..87dc8830
--- /dev/null
+++ b/internal/mcp/init_retry_test.go
@@ -0,0 +1,49 @@
+package mcp
+
+import (
+	"context"
+	"testing"
+	"time"
+)
+
+func TestConnectAndDiscoverRetriesOnStdioInitFailure(t *testing.T) {
+	// "cat" starts instantly but doesn't speak JSON-RPC → Initialize fails every time.
+	// The retry loop should add observable backoff delay before exhausting attempts.
+	ctx, cancel := context.WithTimeout(context.Background(), 5*time.Second)
+	defer cancel()
+
+	start := time.Now()
+	_, _, err := connectAndDiscover(ctx, "test-retry", "stdio",
+		"cat", nil, nil, "", nil, 2)
+	elapsed := time.Since(start)
+
+	if err == nil {
+		t.Fatal("expected error — cat doesn't speak MCP")
+	}
+
+	// First retry waits 2s → elapsed must be at least 2s proving retry happened.
+	if elapsed < 2*time.Second {
+		t.Errorf("elapsed %v — expected ≥2s proving retries happened", elapsed)
+	}
+	t.Logf("elapsed: %v (retries confirmed)", elapsed)
+}
+
+func TestConnectAndDiscoverNoRetryForSSE(t *testing.T) {
+	// SSE/HTTP transports should fail immediately without retrying.
+	ctx, cancel := context.WithTimeout(context.Background(), 5*time.Second)
+	defer cancel()
+
+	start := time.Now()
+	_, _, err := connectAndDiscover(ctx, "test-no-retry", "sse",
+		"", nil, nil, "http://127.0.0.1:1", nil, 10)
+	elapsed := time.Since(start)
+
+	if err == nil {
+		t.Fatal("expected error for bad SSE URL")
+	}
+
+	// SSE should fail fast (no retries) — well under 2 seconds.
+	if elapsed > 2*time.Second {
+		t.Errorf("elapsed %v — SSE should fail fast without retries", elapsed)
+	}
+}
diff --git a/internal/mcp/manager_connect.go b/internal/mcp/manager_connect.go
index 18937feb..9bb6972a 100644
--- a/internal/mcp/manager_connect.go
+++ b/internal/mcp/manager_connect.go
@@ -38,9 +38,37 @@ func connectAndDiscover(ctx context.Context, name, transportType, command string
 		Version: "1.0.0",
 	}
 
-	if _, err := client.Initialize(ctx, initReq); err != nil {
+	// Retry initialization with exponential backoff for slow-starting stdio servers.
+	// Heavy MCP servers (FastMCP with 80+ tools, OAuth servers) can take 3-5s to start
+	// their stdin read loop. Without retries, Initialize sends JSON-RPC before the
+	// server is ready, gets EOF, and permanently fails. SSE/HTTP transports don't need
+	// this because the HTTP server rejects connections until ready (connection refused).
+	const maxInitAttempts = 4 // backoff: 2s + 4s + 8s = ~14s total before giving up
+	var initErr error
+	for attempt := range maxInitAttempts {
+		if attempt > 0 {
+			backoff := time.Duration(1<<attempt) * time.Second // 2s, 4s, 8s
+			slog.Debug("mcp.init.retry", "server", name, "attempt", attempt+1, "backoff", backoff)
+			select {
+			case <-time.After(backoff):
+			case <-ctx.Done():
+				_ = client.Close()
+				return nil, nil, fmt.Errorf("initialize: context cancelled during retry: %w", ctx.Err())
+			}
+		}
+		if _, err := client.Initialize(ctx, initReq); err == nil {
+			break
+		} else {
+			initErr = err
+			// Non-stdio transports: connection errors are definitive, don't retry.
+			if transportType != "stdio" {
+				break
+			}
+		}
+	}
+	if initErr != nil {
 		_ = client.Close()
-		return nil, nil, fmt.Errorf("initialize: %w", err)
+		return nil, nil, fmt.Errorf("initialize: %w", initErr)
 	}
 
 	toolsResult, err := client.ListTools(ctx, mcpgo.ListToolsRequest{})