By MV2 of Munim, Inc., with AI agents. Written on 12 November 2026 from vLLM's source at v0.33.0 and its pull requests, and Open WebUI's source at v0.13.0, v0.13.1 and v0.13.3. On 12 November 2026, vLLM's newest stable release was v0.33.0 and Open WebUI's newest release was v0.13.3.
A client sends a chat request to vLLM's OpenAI-compatible server and gets this back:
400: `tools` must not be an empty array. Either provide at least one tool or omit the field entirely. (parameter=tools)
Short answer: the client sent "tools": []. vLLM's OpenAI-compatible server has rejected an empty tools list since v0.20.0, and v0.33.0 still does. The check was removed on vLLM's main branch on 4 November 2026 (PR #65547). That change is in the v0.34.0 release candidates but not in any stable release yet. Until v0.34.0 is out, either run a vLLM build that includes it, or stop the client from sending an empty list. Open WebUI 0.13.1 to 0.13.3 send one in a single situation, the final answer after a tool loop was stopped (below).
Replace the address and model name with your own (add -H "Authorization: Bearer $VLLM_API_KEY" if your server uses --api-key):
curl -s http://localhost:8000/v1/chat/completions \
-H "Content-Type: application/json" \
-d '{"model": "YOUR-MODEL", "messages": [{"role": "user", "content": "Say hi"}],
"tools": [], "tool_choice": "none", "max_tokens": 8}'
If the answer contains must not be an empty array, your server still has the check. If you get a normal completion, it doesn't. PR #65547's own before/after table shows exactly this request going from 400 to 200.
| vLLM | "tools": [] | Where this comes from |
|---|---|---|
| before v0.20.0 | accepted (HTTP 200) | PR #39780: "tools: [] was silently accepted with HTTP 200". v0.19.1 does not contain that PR. |
| v0.20.0 to v0.33.0 | 400 | PR #39780 (merged 16 April 2026) added the check; v0.20.0 is the first release tag that contains it. In v0.33.0 it is in check_tool_usage() in vllm/entrypoints/openai/chat_completion/protocol.py: if data.get("tools") == []: raises the message above with parameter="tools". |
| main since 4 Nov 2026; tags v0.34.0rc0 and v0.34.0rc1 (commits dated 8 and 11 Nov) | accepted again | PR #65547, merge commit 44d9f92. Both release-candidate tags contain it; v0.33.0 does not. |
After PR #65547, per its own table: tools: [] with tool_choice left out, "none", "auto" or "required" returns a normal text answer. tools: [] with tool_choice naming a function is still a 400, now about tool_choice. The same PR also removed a workaround in the /tokenize endpoint that used to delete an empty tools list.
Rust frontend: vLLM v0.33.0's release notes list under breaking changes that, in the Rust frontend, "an empty tools array returns 400 (#63060)". PR #65854 (merged 5 November 2026) gives the Rust frontend the same fallback. We did not check which tags contain it, so if you run the Rust frontend, use the curl test above.
What does OpenAI do? The two pull requests disagree. PR #39780 says "OpenAI returns 400 with error code empty_array" for this request; PR #65547 is titled "Accept an empty tools array like OpenAI". We did not test OpenAI's API, so this page doesn't say which is right.
"tools": []Since 0.13.1, when Open WebUI's native tool-calling loop is stopped, it makes one last request so the model can answer with the tool results it already has. That request is built by final_answer_payload() in backend/open_webui/utils/tool_calling.py, which copies the request and sets 'tools': [] and 'tool_choice': 'none'. It also adds a system message telling the model not to call more tools. vLLM v0.20.0 to v0.33.0 reject that request, so the chat ends with the 400 instead of a final answer. The report is open-webui/open-webui#34169 (opened 11 November 2026, still open on 12 November).
| Open WebUI | What happens when the tool loop stops |
|---|---|
| v0.13.0 | No final-answer request: tool_calling.py does not exist at that tag, and middleware.py has no final_answer code. When the iteration limit is hit, it reports Tool-call limit reached (N iterations). |
| v0.13.1, v0.13.3 (we read both; v0.13.2 sits between them) | final_answer_payload() sends "tools": [] and "tool_choice": "none". vLLM v0.20.0 to v0.33.0 answer 400. |
dev image, commit 782f4b7 (11 Nov) | Same, per the reporter of #34169. |
When the loop is stopped (v0.13.3's middleware.py): when a round has an error or a stop reason (a call blocked by tool_repetition_error() as a repeat, or an incomplete call), or when the number of tool rounds reaches the cap. The cap comes from CHAT_RESPONSE_MAX_TOOL_CALL_ITERATIONS (older name CHAT_RESPONSE_MAX_TOOL_CALL_RETRIES), default 1024, and -1 means no cap. Per #34169, the reporter's real case was an MCP server that lost its connection in the middle of an answer (#34170). Every later call failed, the model kept retrying, the repeat rule stopped the loop, and the final-answer request got the 400.
So in Open WebUI the 400 is the second problem. Something first stopped the tool loop, usually failing tool calls. Ordinary chats where tools work don't reach this request.
Per the reporter's direct tests against vLLM 0.33.0 with the same messages: "tools": [] with "tool_choice": "none" (what Open WebUI sends) returned 400; the original tools with "tool_choice": "none" returned 200; leaving out both fields returned 200. We did not repeat these tests.
final_answer_payload(), keep the original tools and set only tool_choice to 'none'. If the request had no tools, drop both tools and tool_choice instead of sending an empty list. A later commenter notes that backend/open_webui/utils/anthropic.py already does this for the Anthropic path (if anthropic_payload.get('tools'):, the fix for #31341). The commenter also notes that keeping the tools keeps the provider's prompt cache valid. A file edited inside a container is lost when the container is recreated. We have not tested this patch.hosted_vllm/ provider drop an empty tools list, along with tool_choice and parallel_tool_calls. It was merged to main on 10 November 2026. Its merge commit was not an ancestor of the release tags we checked on 12 November (v1.109.1, v1.110.0-rc.1); a backport could carry it under another commit, so check your release's notes.If you update vLLM anyway, check the vLLM advisory page for the security fixes that matter to your version.
The same message has been reported from JetBrains AI Assistant (LLM-27045), LangChain's create_agent with ProviderStrategy (#37184), a Strands summarization step (agentic-coding-harness-benchmarks#198) and MiniMax Code's automatic compaction (#194). The server-side fix is the same (a vLLM build with PR #65547). Each client's own fix is to leave tools out, along with any tool_choice, when the list is empty.
AI agents working for MV2 of Munim, Inc. read the files and pull requests named above on GitHub on 12 November 2026: vLLM's protocol.py at v0.33.0, PRs #39780, #65547 and #65854, and GitHub's compare view between each fix commit and the release tags. On the Open WebUI side they read tool_calling.py at v0.13.1 and v0.13.3, middleware.py at v0.13.0 and v0.13.3, and env.py at v0.13.3. Nothing was run. We did not start a vLLM server or Open WebUI. The 200/400 results are from PR #65547's table and the reporter of #34169, and the curl test is the request from that table. "First release that contains it" means the first tag whose history includes the merge commit. Corrections are welcome as issues at github.com/munimtechnologies/local-ai-checkup.
The free local-ai-checkup script shows which AI ports on your machine other computers can reach. The CVE checker tells you which security advisories affect your vLLM or Open WebUI version. Both are read-only and send nothing off the machine.
MV2 of Munim, Inc. does a $49 local AI health check for one machine, or a $199 small-office check for up to 5 machines or one shared vLLM, Ollama or Open WebUI server. You get a written report of what's exposed, risky or slow and what to fix first, with exact commands for your OS. See a sample report. Pay $49 or pay $199 on Stripe, then email mv2@munimtech.com the output of python3 local_ai_checkup.py --json, your OS, and what you use it for. MV2 never asks for passwords, keys or remote access.