Problem#
serverTools is null on 66 chat rows, and none of them has a factSources.serverTools entry. No row has []. A reader cannot tell "this model has no built-in tools" from "not filled yet".
#409 asked for this and is closed, but its acceptance is not met. Snapshot 2026-11-21, GET /v1/models?provider=…&provenance=1, chat rows only.
| Provider | Chat rows | serverTools null, no factSources entry | Examples |
|---|
| openai | 70 | 19 | gpt-5-search-api, gpt-audio, gpt-audio-1.5 |
| mistral | 32 | 16 | codestral-2508, ministral-14b-2512, glm-5-2 |
| gemini | 31 | 12 | antigravity-preview-09-2026, aqa |
| byteplus | 23 | 11 | deepseek-v3-2-251201, glm-4-7-251222, seed-1-6-250615 |
| groq | 7 | 5 | allam-2-7b, openai/gpt-oss-safeguard-20b |
| grok | 7 | 3 | grok-4.3, grok-4.20-multi-agent-0309 |
Two lists also look short:
- grok: every filled row has only
["attachment_search"]. #409 says xAI documents web_search and x_search.
- gemini: the vocabulary across all rows is
codeExecution, computerUse, fileSearch, googleMaps, urlContext. Grounding with Google Search is not in it.
Impact#
TanStack AI writes each new model's built-in tool list from this field, with no second source and no hand-written list. A null with no entry forces a guess.
Ask#
- When the provider documents no built-in tools for a model, return
serverTools: [], or keep null and add factSources.serverTools with derivation: "source-silent" (as #273 does for maxOutput).
- Check the Grok and Gemini tool lists against the provider docs.
Acceptance#
- Every chat row of the six providers above has
serverTools as an array, or a factSources.serverTools entry.
#721 is merged. It does not close this issue.
What landed:
factSources.serverToolswithderivation: "source-silent"on 27 chat rows where the provider's per-model source was read and states no built-in tool (mistral 16, openai 7, gemini 3, byteplus 1), the Gemini "Grounding with Google Search" label is read again, andgemini-flash-latestcopies its stated alias target's tools.Still neither an array nor an entry: 38 chat rows. Their provider does not state tools per model on the page we read, so calling them source-silent would be false.
gpt-realtime*(10 ids),gpt-live-1,gpt-live-1.1antigravity-preview-05-2026,-09-2026,-latest,deep-research-preview-04-2026,deep-research-max-preview-04-2026,gemini-2.5-computer-use-preview-10-2025,gemma-4-26b-a4b-it,gemma-4-31b-itdeepseek-v3-2-251201,glm-4-7-251222,gpt-oss-120b-250805,seed-1-6-*(4),seed-1-8-251228,seed-2-0-lite-260228,seed-2-0-mini-260215allam-2-7b,meta-llama/llama-prompt-guard-2-22m,-86m,openai/gpt-oss-safeguard-20b,qwen/qwen3.8-27bgrok-4.20-0309-non-reasoning,grok-4.20-multi-agent-0309,grok-4.3Sources a parser could read next: Groq's browser-search and code-execution "Supported Models" tables; Gemini's antigravity-agent, deep-research and computer-use guides and the Gemma on Gemini API guide; OpenAI's realtime MCP guide; xAI's multi-agent page (it names doc names, not request type ids, so it needs a mapping the page itself states).