Misc. bug: Server UI HTTP 404 on /tools

[rchildre3]

Name and Version

version: 9755 (0ef6f06d5)
built with GNU 15.2.0 for Linux x86_64

Built with:

cmake \
  -G Ninja \
  -B build \
  -DCMAKE_BUILD_TYPE=RelWithDebInfo \
  -DCMAKE_CXX_FLAGS="-march=native" \
  -DCMAKE_EXPORT_COMPILE_COMMANDS=1 \
  -DCMAKE_INTERPROCEDURAL_OPTIMIZATION=1 \
  -DGGML_VULKAN=1 \
  -DGGML_NATIVE=1

Operating systems

Linux

Which llama.cpp modules do you know to be affected?

llama-server

Command line

pwd 
$HOME/llama.cpp

ls -CF ./tools/
CMakeLists.txt      export-lora/  mtmd/        rpc/
batched-bench/      fit-params/   parser/      server/
cli/                gguf-split/   perplexity/  tokenize/
completion/         imatrix/      quantize/    tts/
cvector-generator/  llama-bench/  results/     ui/

ls -CF ./build/tools/
CMakeFiles/          completion/         imatrix/      quantize/  ui/
CTestTestfile.cmake  cvector-generator/  llama-bench/  results/
batched-bench/       export-lora/        mtmd/         server/
cli/                 fit-params/         parser/       tokenize/
cmake_install.cmake  gguf-split/         perplexity/   tts/

./build/bin/llama-server -hf unsloth/gemma-4-E2B-it-qat-GGUF:UD-Q4_K_XL --threads 4 --n-gpu-layers 0

Problem description & steps to reproduce

in browser, at http://127.0.0.1:8080/

DevTools Console shows:

XHRGET http://127.0.0.1:8080/tools [HTTP/1.1 404 Not Found 0ms]

First Bad Commit

No response

Relevant log output

Logs

[macrosoft]

I am running the llama server on my local network.

Initially, this 404 error appeared yesterday on one of my machines, but my main computer was running correctly. However, today after updating llama.cpp, the issue has manifested on all computers.

Observed Error:
curl http://myserver:8080/tools
{"error":{"message":"File Not Found","type":"not_found_error","code":404}}

Note that rolling back to the previous version of llama.cpp did not resolve the issue either

[karan68]

I can reproduce this on the current master (build 9758) with a CPU-only server.

llama-server -m model.gguf --host 127.0.0.1 --port 8080 -ngl 0

curl http://127.0.0.1:8080/tools
{"error":{"message":"File Not Found","type":"not_found_error","code":404}}

/props, /, and /v1/models all return 200, only /tools 404s.

Root cause: in server.cpp, the /tools GET/POST routes are only registered when the server is started with built-in tools enabled:

// EXPERIMENTAL built-in tools
if (!params.server_tools.empty()) {
    ...
    ctx_http.get ("/tools", ex_wrapper(tools.handle_get));
    ctx_http.post("/tools", ex_wrapper(tools.handle_post));
}

So without --tools, the route doesn't exist and the WebUI's unconditional GET /tools on load (api-endpoints.ts TOOLS.LIST = '/tools', called from tools.svelte.ts fetchBuiltinTools()) gets a 404. The WebUI already treats this 404 as "no built-in tools" and keeps working, so this is cosmetic console noise rather than a functional break, but it does log an error on every page load.

Proposed fix: always register GET /tools and have it return an empty list [] when no tools are enabled, keeping POST /tools and the experimental security warnings gated behind tools actually being enabled. handle_get already returns json::array() over the (empty) tool list, so this is a small change in server.cpp.

Happy to open a PR for this if it sounds reasonable.