diff --git a/clients/clients.go b/clients/clients.go
index 2cb14c7..59cab82 100644
--- a/clients/clients.go
+++ b/clients/clients.go
@@ -6,6 +6,7 @@
 package clients
 
 import (
+	"bufio"
 	"bytes"
 	"context"
 	"encoding/json"
@@ -14,6 +15,7 @@ import (
 	"mime/multipart"
 	"net/http"
 	"net/url"
+	"strings"
 	"sync"
 	"time"
 )
@@ -142,6 +144,36 @@ func (h *httpClient) do(req *http.Request) ([]byte, error) {
 	return body, nil
 }
 
+// postJSONStream sends a JSON POST and returns the raw *http.Response so the
+// caller can read the body incrementally (e.g. for SSE streaming). The caller
+// is responsible for closing resp.Body.
+func (h *httpClient) postJSONStream(ctx context.Context, path string, body any) (*http.Response, error) {
+	buf := getBuf()
+	defer putBuf(buf)
+	if err := json.NewEncoder(buf).Encode(body); err != nil {
+		return nil, fmt.Errorf("marshal: %w", err)
+	}
+	// Copy to a non-pooled buffer so we can safely return the pool buffer.
+	payload := make([]byte, buf.Len())
+	copy(payload, buf.Bytes())
+
+	req, err := http.NewRequestWithContext(ctx, http.MethodPost, h.baseURL+path, bytes.NewReader(payload))
+	if err != nil {
+		return nil, err
+	}
+	req.Header.Set("Content-Type", "application/json")
+	resp, err := h.client.Do(req)
+	if err != nil {
+		return nil, fmt.Errorf("http %s %s: %w", req.Method, req.URL.Path, err)
+	}
+	if resp.StatusCode >= 400 {
+		respBody, _ := io.ReadAll(resp.Body)
+		_ = resp.Body.Close()
+		return nil, fmt.Errorf("http %d: %s", resp.StatusCode, string(respBody))
+	}
+	return resp, nil
+}
+
 func (h *httpClient) healthCheck(ctx context.Context) bool {
 	data, err := h.get(ctx, "/health", nil)
 	_ = data
@@ -320,6 +352,73 @@ func (c *LLMClient) Generate(ctx context.Context, prompt string, context_ string
 	return resp.Choices[0].Message.Content, nil
 }
 
+// StreamGenerate sends a streaming chat completion request and calls onToken
+// for each content delta received via SSE. Returns the fully assembled text.
+// The onToken callback is invoked synchronously on the calling goroutine; it
+// should be fast (e.g. publish a NATS message).
+func (c *LLMClient) StreamGenerate(ctx context.Context, prompt string, context_ string, systemPrompt string, onToken func(token string)) (string, error) {
+	msgs := buildMessages(prompt, context_, systemPrompt)
+	payload := map[string]any{
+		"model":       c.Model,
+		"messages":    msgs,
+		"max_tokens":  c.MaxTokens,
+		"temperature": c.Temperature,
+		"top_p":       c.TopP,
+		"stream":      true,
+	}
+
+	resp, err := c.postJSONStream(ctx, "/v1/chat/completions", payload)
+	if err != nil {
+		return "", err
+	}
+	defer func() { _ = resp.Body.Close() }()
+
+	var full strings.Builder
+	scanner := bufio.NewScanner(resp.Body)
+	// SSE lines can be up to 64 KiB for large token batches.
+	scanner.Buffer(make([]byte, 0, 64*1024), 64*1024)
+
+	for scanner.Scan() {
+		line := scanner.Text()
+		if !strings.HasPrefix(line, "data: ") {
+			continue
+		}
+		data := strings.TrimPrefix(line, "data: ")
+		if data == "[DONE]" {
+			break
+		}
+		var chunk struct {
+			Choices []struct {
+				Delta struct {
+					Content string `json:"content"`
+				} `json:"delta"`
+			} `json:"choices"`
+		}
+		if err := json.Unmarshal([]byte(data), &chunk); err != nil {
+			continue // skip malformed chunks
+		}
+		if len(chunk.Choices) == 0 {
+			continue
+		}
+		token := chunk.Choices[0].Delta.Content
+		if token == "" {
+			continue
+		}
+		full.WriteString(token)
+		if onToken != nil {
+			onToken(token)
+		}
+	}
+	if err := scanner.Err(); err != nil {
+		// If we already collected some text, return it with the error.
+		if full.Len() > 0 {
+			return full.String(), fmt.Errorf("stream interrupted: %w", err)
+		}
+		return "", fmt.Errorf("stream read: %w", err)
+	}
+	return full.String(), nil
+}
+
 func buildMessages(prompt, ctx, systemPrompt string) []ChatMessage {
 	var msgs []ChatMessage
 	if systemPrompt != "" {
diff --git a/clients/clients_test.go b/clients/clients_test.go
index 0406552..9c24887 100644
--- a/clients/clients_test.go
+++ b/clients/clients_test.go
@@ -4,6 +4,7 @@ import (
 	"bytes"
 	"context"
 	"encoding/json"
+	"fmt"
 	"io"
 	"net/http"
 	"net/http/httptest"
@@ -299,6 +300,293 @@ func TestLLMClient_GenerateNoChoices(t *testing.T) {
 	}
 }
 
+// ────────────────────────────────────────────────────────────────────────────
+// LLM client — StreamGenerate
+// ────────────────────────────────────────────────────────────────────────────
+
+// sseChunk builds an OpenAI-compatible SSE chat.completion.chunk line.
+func sseChunk(content string) string {
+	chunk := map[string]any{
+		"choices": []map[string]any{
+			{"delta": map[string]any{"content": content}},
+		},
+	}
+	b, _ := json.Marshal(chunk)
+	return "data: " + string(b) + "\n\n"
+}
+
+func TestLLMClient_StreamGenerate(t *testing.T) {
+	tokens := []string{"Hello", " world", "!"}
+	ts := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		if r.URL.Path != "/v1/chat/completions" {
+			t.Errorf("path = %q", r.URL.Path)
+		}
+		var req map[string]any
+		_ = json.NewDecoder(r.Body).Decode(&req)
+		if req["stream"] != true {
+			t.Errorf("stream = %v, want true", req["stream"])
+		}
+
+		w.Header().Set("Content-Type", "text/event-stream")
+		flusher, _ := w.(http.Flusher)
+		for _, tok := range tokens {
+			_, _ = w.Write([]byte(sseChunk(tok)))
+			flusher.Flush()
+		}
+		_, _ = w.Write([]byte("data: [DONE]\n\n"))
+		flusher.Flush()
+	}))
+	defer ts.Close()
+
+	c := NewLLMClient(ts.URL, 5*time.Second)
+	var received []string
+	result, err := c.StreamGenerate(context.Background(), "hi", "", "", func(tok string) {
+		received = append(received, tok)
+	})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if result != "Hello world!" {
+		t.Errorf("result = %q, want %q", result, "Hello world!")
+	}
+	if len(received) != 3 {
+		t.Fatalf("callback count = %d, want 3", len(received))
+	}
+	if received[0] != "Hello" || received[1] != " world" || received[2] != "!" {
+		t.Errorf("received = %v", received)
+	}
+}
+
+func TestLLMClient_StreamGenerateWithSystemPrompt(t *testing.T) {
+	ts := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		var req map[string]any
+		_ = json.NewDecoder(r.Body).Decode(&req)
+		msgs, _ := req["messages"].([]any)
+		if len(msgs) != 2 {
+			t.Errorf("expected system+user, got %d messages", len(msgs))
+		}
+		first, _ := msgs[0].(map[string]any)
+		if first["role"] != "system" || first["content"] != "You are a DM" {
+			t.Errorf("system msg = %v", first)
+		}
+
+		w.Header().Set("Content-Type", "text/event-stream")
+		flusher, _ := w.(http.Flusher)
+		_, _ = w.Write([]byte(sseChunk("ok")))
+		_, _ = w.Write([]byte("data: [DONE]\n\n"))
+		flusher.Flush()
+	}))
+	defer ts.Close()
+
+	c := NewLLMClient(ts.URL, 5*time.Second)
+	result, err := c.StreamGenerate(context.Background(), "roll dice", "", "You are a DM", nil)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if result != "ok" {
+		t.Errorf("result = %q", result)
+	}
+}
+
+func TestLLMClient_StreamGenerateNilCallback(t *testing.T) {
+	ts := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		w.Header().Set("Content-Type", "text/event-stream")
+		flusher, _ := w.(http.Flusher)
+		_, _ = w.Write([]byte(sseChunk("token")))
+		_, _ = w.Write([]byte("data: [DONE]\n\n"))
+		flusher.Flush()
+	}))
+	defer ts.Close()
+
+	c := NewLLMClient(ts.URL, 5*time.Second)
+	// nil callback should not panic
+	result, err := c.StreamGenerate(context.Background(), "hi", "", "", nil)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if result != "token" {
+		t.Errorf("result = %q", result)
+	}
+}
+
+func TestLLMClient_StreamGenerateEmptyDelta(t *testing.T) {
+	// SSE chunks with empty content should be silently skipped.
+	ts := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		w.Header().Set("Content-Type", "text/event-stream")
+		flusher, _ := w.(http.Flusher)
+		// role-only chunk (no content) — common for first chunk from vLLM
+		_, _ = w.Write([]byte("data: {\"choices\":[{\"delta\":{\"role\":\"assistant\"}}]}\n\n"))
+		// empty content string
+		_, _ = w.Write([]byte(sseChunk("")))
+		// real token
+		_, _ = w.Write([]byte(sseChunk("hello")))
+		_, _ = w.Write([]byte("data: [DONE]\n\n"))
+		flusher.Flush()
+	}))
+	defer ts.Close()
+
+	c := NewLLMClient(ts.URL, 5*time.Second)
+	var count int
+	result, err := c.StreamGenerate(context.Background(), "q", "", "", func(tok string) {
+		count++
+	})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if result != "hello" {
+		t.Errorf("result = %q", result)
+	}
+	if count != 1 {
+		t.Errorf("callback count = %d, want 1 (empty deltas should be skipped)", count)
+	}
+}
+
+func TestLLMClient_StreamGenerateMalformedChunks(t *testing.T) {
+	// Malformed JSON should be skipped without error.
+	ts := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		w.Header().Set("Content-Type", "text/event-stream")
+		flusher, _ := w.(http.Flusher)
+		_, _ = w.Write([]byte("data: {invalid json}\n\n"))
+		_, _ = w.Write([]byte("data: {\"choices\":[]}\n\n")) // empty choices
+		_, _ = w.Write([]byte(sseChunk("good")))
+		_, _ = w.Write([]byte("data: [DONE]\n\n"))
+		flusher.Flush()
+	}))
+	defer ts.Close()
+
+	c := NewLLMClient(ts.URL, 5*time.Second)
+	result, err := c.StreamGenerate(context.Background(), "q", "", "", nil)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if result != "good" {
+		t.Errorf("result = %q, want %q", result, "good")
+	}
+}
+
+func TestLLMClient_StreamGenerateHTTPError(t *testing.T) {
+	ts := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		w.WriteHeader(500)
+		_, _ = w.Write([]byte("internal server error"))
+	}))
+	defer ts.Close()
+
+	c := NewLLMClient(ts.URL, 5*time.Second)
+	_, err := c.StreamGenerate(context.Background(), "q", "", "", nil)
+	if err == nil {
+		t.Fatal("expected error for 500")
+	}
+	if !strings.Contains(err.Error(), "500") {
+		t.Errorf("error should contain 500: %v", err)
+	}
+}
+
+func TestLLMClient_StreamGenerateContextCanceled(t *testing.T) {
+	started := make(chan struct{})
+	ts := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		w.Header().Set("Content-Type", "text/event-stream")
+		flusher, _ := w.(http.Flusher)
+		// Send several tokens so the client receives some before cancel.
+		for i := range 20 {
+			_, _ = w.Write([]byte(sseChunk(fmt.Sprintf("tok%d ", i))))
+			flusher.Flush()
+		}
+		close(started)
+		// Block until client cancels
+		<-r.Context().Done()
+	}))
+	defer ts.Close()
+
+	ctx, cancel := context.WithCancel(context.Background())
+	c := NewLLMClient(ts.URL, 10*time.Second)
+
+	var streamErr error
+	done := make(chan struct{})
+	go func() {
+		defer close(done)
+		_, streamErr = c.StreamGenerate(ctx, "q", "", "", nil)
+	}()
+
+	<-started
+	cancel()
+	<-done
+
+	// After cancel the stream should return an error (context canceled or
+	// stream interrupted). The exact partial text depends on timing.
+	if streamErr == nil {
+		t.Error("expected error after context cancel")
+	}
+}
+
+func TestLLMClient_StreamGenerateNoSSEPrefix(t *testing.T) {
+	// Lines without "data: " prefix should be silently ignored (comments, blank lines, event IDs).
+	ts := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		w.Header().Set("Content-Type", "text/event-stream")
+		flusher, _ := w.(http.Flusher)
+		_, _ = w.Write([]byte(": this is an SSE comment\n\n"))
+		_, _ = w.Write([]byte("event: message\n"))
+		_, _ = w.Write([]byte(sseChunk("word")))
+		_, _ = w.Write([]byte("\n")) // blank line
+		_, _ = w.Write([]byte("data: [DONE]\n\n"))
+		flusher.Flush()
+	}))
+	defer ts.Close()
+
+	c := NewLLMClient(ts.URL, 5*time.Second)
+	result, err := c.StreamGenerate(context.Background(), "q", "", "", nil)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if result != "word" {
+		t.Errorf("result = %q, want %q", result, "word")
+	}
+}
+
+func TestLLMClient_StreamGenerateManyTokens(t *testing.T) {
+	// Verify token ordering and full assembly with many chunks.
+	n := 100
+	ts := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		w.Header().Set("Content-Type", "text/event-stream")
+		flusher, _ := w.(http.Flusher)
+		for i := range n {
+			tok := fmt.Sprintf("t%d ", i)
+			_, _ = w.Write([]byte(sseChunk(tok)))
+			flusher.Flush()
+		}
+		_, _ = w.Write([]byte("data: [DONE]\n\n"))
+		flusher.Flush()
+	}))
+	defer ts.Close()
+
+	c := NewLLMClient(ts.URL, 5*time.Second)
+	var mu sync.Mutex
+	var order []int
+	result, err := c.StreamGenerate(context.Background(), "q", "", "", func(tok string) {
+		var idx int
+		fmt.Sscanf(tok, "t%d ", &idx)
+		mu.Lock()
+		order = append(order, idx)
+		mu.Unlock()
+	})
+	if err != nil {
+		t.Fatal(err)
+	}
+	// Verify all tokens arrived in order
+	if len(order) != n {
+		t.Fatalf("got %d tokens, want %d", len(order), n)
+	}
+	for i, v := range order {
+		if v != i {
+			t.Errorf("order[%d] = %d", i, v)
+			break
+		}
+	}
+	// Quick sanity: result should start with "t0 " and end with last token
+	if !strings.HasPrefix(result, "t0 ") {
+		t.Errorf("result prefix = %q", result[:10])
+	}
+}
+
 // ────────────────────────────────────────────────────────────────────────────
 // TTS client
 // ────────────────────────────────────────────────────────────────────────────