diff --git a/README.md b/README.md
index 8318527..462fd6c 100644
--- a/README.md
+++ b/README.md
@@ -174,6 +174,17 @@ Selected with `--profile`:
 The server is OpenAI-compatible (`/v1/chat/completions` with tool calls + SSE,
 `/v1/completions`, `/v1/models`) and serves under the model name `qwen`.
 
+**Tool calling is enabled by default** (`--enable-auto-tool-choice
+--tool-call-parser qwen3_xml --reasoning-parser qwen3`), so clients can send `tools`
+with `tool_choice="auto"`. Qwen3.5 emits the *XML* tool format, so `qwen3_xml` is the
+correct parser — `hermes` returns 200 but with empty `tool_calls` (the call lands in
+`content` instead). `<think>` reasoning is split into `reasoning_content`. Override
+with `TOOL_PARSER=` / `REASONING_PARSER=` (empty disables; `qwen3_coder` also works).
+
+**Network:** the server binds `0.0.0.0` in `--net=host`, so it is reachable from the
+LAN at `http://<spark-ip>:8000` — the `127.0.0.1` in the examples is just the local
+default. There is no auth; put it behind a reverse proxy / firewall for shared use.
+
 ## Benchmarks
 
 All single-stream (c=1), temperature 0, GB10. "Hermes" regenerates the next
diff --git a/runtime/mtp_serve.sh b/runtime/mtp_serve.sh
index 5223378..3112c3b 100755
--- a/runtime/mtp_serve.sh
+++ b/runtime/mtp_serve.sh
@@ -15,6 +15,12 @@ LOAD_FORMAT="${LOAD_FORMAT:-fastsafetensors}"
 PORT="${PORT:-8000}"
 # Reclaim the CUDA-graph memory over-estimate to KV (see serve.sh). Set =1 to restore.
 export VLLM_MEMORY_PROFILER_ESTIMATE_CUDAGRAPHS="${VLLM_MEMORY_PROFILER_ESTIMATE_CUDAGRAPHS:-0}"
+# Auto tool-calling + reasoning split (see serve.sh). Off via TOOL_PARSER="" etc.
+TOOL_PARSER="${TOOL_PARSER:-qwen3_xml}"
+REASONING_PARSER="${REASONING_PARSER:-qwen3}"
+TOOL_ARG=()
+[ -n "$TOOL_PARSER" ] && TOOL_ARG+=(--enable-auto-tool-choice --tool-call-parser "$TOOL_PARSER")
+[ -n "$REASONING_PARSER" ] && TOOL_ARG+=(--reasoning-parser "$REASONING_PARSER")
 echo "[mtp] qwen3_5_mtp — backend=$BACKEND, num_speculative_tokens=$NSPEC, model=$MODEL"
 exec vllm serve "$MODEL" \
   --served-model-name qwen \
@@ -28,4 +34,5 @@ exec vllm serve "$MODEL" \
   --trust-remote-code \
   --load-format "$LOAD_FORMAT" \
   --attention-backend "$BACKEND" \
+  "${TOOL_ARG[@]}" \
   --speculative-config "{\"method\":\"qwen3_5_mtp\",\"num_speculative_tokens\":$NSPEC,\"model\":\"$MODEL\"}"
diff --git a/runtime/serve.sh b/runtime/serve.sh
index 05d6c57..63333ee 100755
--- a/runtime/serve.sh
+++ b/runtime/serve.sh
@@ -60,6 +60,22 @@ else
 fi
 python3 /host/patch_unify2.py || { [ "$NSPEC" = "0" ] && true; }
 
+# OpenAI automatic tool-calling + reasoning split. Without --enable-auto-tool-choice
+# + --tool-call-parser, any client that sends `tools` with tool_choice="auto" gets
+# HTTP 400 — so agent/tool workloads (the whole point here) need these on. The tool
+# parser only activates when a request carries `tools`, so leaving it on is free for
+# plain chat. Qwen3.5 emits the XML tool format
+# (<tool_call><function=name><parameter=k>v</parameter></function></tool_call>) -> the
+# qwen3_xml parser (NOT hermes, which only reads the JSON format -> empty tool_calls),
+# and <think> reasoning the qwen3 parser splits into `reasoning_content`. Verified:
+# tool_choice="auto" -> tool_calls=[get_weather {"city":"Paris"}]. Disable either with
+# TOOL_PARSER="" / REASONING_PARSER="", or override the name (e.g. qwen3_coder).
+TOOL_PARSER="${TOOL_PARSER:-qwen3_xml}"
+REASONING_PARSER="${REASONING_PARSER:-qwen3}"
+TOOL_ARG=()
+[ -n "$TOOL_PARSER" ] && TOOL_ARG+=(--enable-auto-tool-choice --tool-call-parser "$TOOL_PARSER")
+[ -n "$REASONING_PARSER" ] && TOOL_ARG+=(--reasoning-parser "$REASONING_PARSER")
+
 exec vllm serve "$MODEL" \
   --served-model-name qwen \
   --host 0.0.0.0 --port "$PORT" \
@@ -72,4 +88,5 @@ exec vllm serve "$MODEL" \
   --trust-remote-code \
   --load-format "$LOAD_FORMAT" \
   --attention-backend "$BACKEND" \
+  "${TOOL_ARG[@]}" \
   "${SPEC_ARG[@]}"