Tool execution — OpenAI Responses · Anthropic Messages · xAI Responses · Gemini generateContent / Interactions
Status: shapes and statuses from generated/tools.json (69 records: OpenAI 19, Anthropic 27, xAI 12, Gemini 11), generated/parameters.json (tool entries), generated/compatibility/{model-tool-matrix,gemini-tool-model-matrix}.json; live observations quoted from docs/openai/tool-loop.md, docs/tools/openai/.md, docs/tools/anthropic/.md, docs/tools/xai/.md, docs/xai/tool-loop.md, docs/tools/gemini/.md, docs/gemini/tool-loop.md (2026-09-18/19). Gemini googleSearch and computerUse are ACCOUNT_RESTRICTED on this free-tier key; xAI tool_search is alpha-only (403).
Sources: https://developers.openai.com/api/docs/guides/function-calling · …/guides/tools · https://platform.claude.com/docs/en/agents-and-tools/tool-use/overview · https://docs.x.ai/developers/tools/overview · https://docs.x.ai/developers/model-capabilities/text/function-calling · https://ai.google.dev/gemini-api/docs/function-calling · https://ai.google.dev/gemini-api/docs/tools
Last verified: 2026-09-18
1. Taxonomy
| Category | OpenAI (tools[].type) |
Anthropic (tools[].type) |
xAI (tools[].type, Responses) |
Gemini (tools[] keys) |
Who executes |
|---|---|---|---|---|---|
| Your own tools | function, custom (free-form / grammar), grouped in namespace |
custom tool (name + input_schema) |
function (also the only tool type on Chat Completions and /v1/messages) |
functionDeclarations[] (Interactions {type: function}) |
you |
| Vendor-defined, you execute | computer, computer_use_preview, apply_patch, shell (environment.type: local) |
bash_20250124, text_editor_20250728, memory_20250818, computer_toolset_20260801, computer_20251124, browser_toolset_20260801 |
shell (environment: {type: local}) |
computerUse (predefined functionCalls: click, type, scroll, navigate, open_app…; PREVIEW) |
you |
| Vendor-executed (hosted / server) | web_search, file_search, code_interpreter, image_generation, shell (container), tool_search (server), programmatic_tool_calling |
web_search_*, web_fetch_*, code_execution_*, tool_search_*, advisor_20260301 |
web_search, x_search, code_interpreter (= code_execution), file_search (= collections_search), image_generation, implicit attachment search, view_image/view_x_video sub-tools, tool_search (alpha) |
googleSearch (+ searchTypes.imageSearch), googleMaps, urlContext, codeExecution, fileSearch; legacy googleSearchRetrieval |
vendor, inside the request |
| Remote MCP | mcp {server_url|connector_id|tunnel_id} |
mcp_servers[] + mcp_toolset (beta) |
mcp {server_url, server_label, allowed_tools, authorization, headers} |
mcpServers[] {streamableHttpTransport} (schema only, UNVERIFIED) / Interactions {type: mcp_server}; SDK-side MCP (mcpToTool()) |
vendor calls the server (Gemini SDK-side: your process) |
| Skills | skill_reference in shell.environment.skills[], containers, Agents |
container.skills[], Managed Agents skills[] |
shell.environment.skills[] {name, description, path} (Skills API 404 for this team) |
.agents/skills/<name>/SKILL.md in custom-agent environments |
vendor sandbox / local |
2. Definition mapping (client tools)
| Field | OpenAI tools[type=function] |
Anthropic custom tool | xAI tools[type=function] |
Gemini functionDeclarations[] |
Notes |
|---|---|---|---|---|---|
| name | name |
name (^[a-zA-Z0-9_-]{1,128}$) |
name |
name (≤128, [A-Za-z0-9_.:-], start with letter/underscore) |
|
| description | description |
description |
description |
description (required per REST reference) |
|
| schema | parameters (JSON Schema) |
input_schema (JSON Schema) |
parameters (JSON Schema; root must be an object) |
parameters (OpenAPI 3.03 subset) or parametersJsonSchema (full JSON Schema) |
|
| strict | strict: true (omit → try strict, fall back) |
strict: true (≤20 strict tools) |
always strict (strict accepted, ignored) |
toolConfig.functionCallingConfig.mode: VALIDATED (request-level) |
|
| examples | — | input_examples[] |
— | — | |
| result schema | output_schema (programmatic callers) |
— | — | response / responseJsonSchema |
|
| lazy load | defer_loading: true (+ tool_search) |
defer_loading: true (+ tool search tool) |
defer_loading: true (+ tool_search, alpha 403) |
— (best practice 10–20 active declarations; max 512) | |
| callers | allowed_callers: ["direct","programmatic"] |
allowed_callers: ["direct","code_execution_20260120"] |
— | — | |
| async | async: true (GPT-6 Astra) |
— | — | behavior: NON_BLOCKING (Live API only) |
|
| cache breakpoint | per-part prompt_cache_breakpoint (5.6+) |
cache_control on the tool |
automatic | inside cachedContents.tools |
|
| limits | 2,000 per agent | 128 per agent | ≤350 tools (Chat) | 512 declarations (SDK) |
3. The loop — identical task on all four providers
Task: the model must call get_weather(city); you run it and return the result.
Request 1
json
// OpenAI — POST /v1/responses
{"model": "gpt-5.4-nano", "input": "Weather in Paris?",
"tools": [{"type": "function", "name": "get_weather", "description": "Current weather for a city",
"parameters": {"type": "object", "properties": {"city": {"type": "string"}}, "required": ["city"], "additionalProperties": false}, "strict": true}],
"tool_choice": "auto", "parallel_tool_calls": true}json
// Anthropic — POST /v1/messages
{"model": "claude-haiku-4-5-20251001", "max_tokens": 256,
"messages": [{"role": "user", "content": "Weather in Paris?"}],
"tools": [{"name": "get_weather", "description": "Current weather for a city",
"input_schema": {"type": "object", "properties": {"city": {"type": "string"}}, "required": ["city"], "additionalProperties": false}, "strict": true}],
"tool_choice": {"type": "auto"}}json
// xAI — POST /v1/responses (same body as OpenAI; strict is implicit)
{"model": "grok-4.3", "input": "Weather in Paris?", "reasoning": {"effort": "low"},
"tools": [{"type": "function", "name": "get_weather", "description": "Current weather for a city",
"parameters": {"type": "object", "properties": {"city": {"type": "string"}}, "required": ["city"]}}],
"tool_choice": "auto", "parallel_tool_calls": true}json
// Gemini — POST /v1beta/models/gemini-3.5-flash-lite:generateContent
{"contents": [{"role": "user", "parts": [{"text": "Weather in Paris?"}]}],
"tools": [{"functionDeclarations": [{"name": "get_weather", "description": "Current weather for a city",
"parameters": {"type": "OBJECT", "properties": {"city": {"type": "STRING"}}, "required": ["city"]}}]}],
"toolConfig": {"functionCallingConfig": {"mode": "AUTO"}}}Response 1 (shapes observed live)
json
// OpenAI output[] item
{"type": "function_call", "id": "fc_…", "call_id": "call_…", "name": "get_weather", "arguments": "{\"city\":\"Paris\"}", "status": "completed"}json
// Anthropic content[] block, stop_reason: "tool_use"
{"type": "tool_use", "id": "toolu_…", "name": "get_weather", "input": {"city": "Paris"}, "caller": {"type": "direct"}}json
// xAI output[] item (preceded by a reasoning item)
{"type": "function_call", "id": "fc_…", "call_id": "call-<uuid>-0", "name": "get_weather", "arguments": "{\"city\":\"Paris\"}", "status": "completed"}json
// Gemini candidates[0].content.parts[] — the thoughtSignature MUST be echoed back on Gemini 3
{"functionCall": {"name": "get_weather", "args": {"city": "Paris"}, "id": "call_…"}, "thoughtSignature": "<base64>"}Request 2 — return the result
json
// OpenAI (chained)
{"model": "gpt-5.4-nano", "previous_response_id": "resp_…", "tools": [ …same tools… ],
"input": [{"type": "function_call_output", "call_id": "call_…", "output": "{\"temp_c\": 18, \"sky\": \"clear\"}"}]}json
// Anthropic (replay)
{"model": "claude-haiku-4-5-20251001", "max_tokens": 256, "tools": [ …same tools… ],
"messages": [{"role": "user", "content": "Weather in Paris?"},
{"role": "assistant", "content": [{"type": "tool_use", "id": "toolu_…", "name": "get_weather", "input": {"city": "Paris"}}]},
{"role": "user", "content": [{"type": "tool_result", "tool_use_id": "toolu_…", "content": "{\"temp_c\": 18, \"sky\": \"clear\"}"}]}]}json
// xAI (chained — identical to OpenAI; a fresh max_turns budget starts here)
{"model": "grok-4.3", "previous_response_id": "resp_…", "tools": [ …same tools… ],
"input": [{"type": "function_call_output", "call_id": "call-<uuid>-0", "output": "{\"temp_c\": 18, \"sky\": \"clear\"}"}]}json
// Gemini (replay; all functionResponse parts of the step in ONE user Content, after ALL functionCall parts)
{"contents": [{"role": "user", "parts": [{"text": "Weather in Paris?"}]},
{"role": "model", "parts": [{"functionCall": {"name": "get_weather", "args": {"city": "Paris"}, "id": "call_…"}, "thoughtSignature": "<base64>"}]},
{"role": "user", "parts": [{"functionResponse": {"name": "get_weather", "id": "call_…", "response": {"temp_c": 18, "sky": "clear"}}}]}],
"tools": [ …same tools… ]}Differences that break naive ports: arguments is a JSON string on OpenAI/xAI, an object on Anthropic (input) and Gemini (args); the result is a top-level item (OpenAI/xAI), a tool_result block first in a user turn (Anthropic) or a functionResponse part whose response is a JSON object (Gemini); Anthropic signals with stop_reason: tool_use, OpenAI/xAI with a function_call item and status: completed, Gemini with a functionCall part and finishReason: STOP; Gemini 3 rejects a missing/edited thoughtSignature (400 or finishReason: MISSING_THOUGHT_SIGNATURE); tool errors are a string (OpenAI/xAI output), is_error: true (Anthropic) or an {error: …} object in response (Gemini); every call_id / tool_use_id / functionCall.id must be answered.
4. Tool choice and parallelism
| Intent | OpenAI tool_choice |
Anthropic tool_choice |
xAI tool_choice |
Gemini toolConfig.functionCallingConfig (Interactions generation_config.tool_choice) |
|---|---|---|---|---|
| let the model decide | "auto" |
{"type": "auto"} |
"auto" (Chat/Responses); /v1/messages {type: auto} |
{"mode": "AUTO"} ("auto") |
| forbid tools | "none" |
{"type": "none"} |
"none" |
{"mode": "NONE"} ("none") |
| must call some tool | "required" |
{"type": "any"} |
"required" (/v1/messages {type: any}) |
{"mode": "ANY"} (constrained decoding) ("any") |
| must call tool X | {"type": "function", "name": "X"} (also hosted types) |
{"type": "tool", "name": "X"} |
{"type": "function", "name": "X"} (Chat: {type: function, function:{name}}); server tools cannot be forced |
{"mode": "ANY", "allowedFunctionNames": ["X"]}; built-in tools cannot be forced (allowed_tools: {mode, tools[]}) |
| restrict to a subset | {"type": "allowed_tools", "mode": "auto|required", "tools": […]} |
— (remove from tools or defer_loading) |
— | allowedFunctionNames[] (ANY mode) / Interactions allowed_tools |
| schema-validated decoding | strict per tool |
strict per tool |
always | {"mode": "VALIDATED"} (default when built-ins or structured output are combined) ("validated") |
| at most one call | parallel_tool_calls: false |
tool_choice.disable_parallel_tool_use: true |
parallel_tool_calls: false |
— (no switch; the model may emit several functionCall parts) |
| forced choice + reasoning | allowed | any/tool ⟂ manual thinking; 400 on Fable 5.1 / Mythos 5.1 |
allowed (reasoning always on) | allowed; Live setup.toolConfig → close 1007 |
5. Hosted / server tools side by side
| Task | OpenAI | Anthropic | xAI | Gemini | Portable JSON? |
|---|---|---|---|---|---|
| Search the web | {"type":"web_search","search_context_size":"medium","filters":{"allowed_domains":["example.com"]}} → web_search_call + url_citation; $10/1k |
{"type":"web_search_20260318","name":"web_search","max_uses":5,"allowed_domains":["example.com"]} → server_tool_use + web_search_tool_result + web_search_result_location; $10/1k |
{"type":"web_search","allowed_domains":["example.com"],"enable_image_search":true} → web_search_call {action: search|open_page|find_in_page} + url_citation + inline [[N]](url); $5/1k; search_context_size → 400 |
{"googleSearch":{"timeRangeFilter":{…},"searchTypes":{"webSearch":{}}}} → groundingMetadata {webSearchQueries, searchEntryPoint (must render), groundingChunks[].web, groundingSupports}; 5,000 free queries/month then $14/1k (3.x) |
concept yes; shapes no |
| Search social / places | — | — | {"type":"x_search","allowed_x_handles":["xai"],"from_date":"2026-09-01","enable_video_understanding":true} → custom_tool_call (x_keyword_search/x_semantic_search); $5/1k → per-post/profile from 2026-09-21 |
{"googleMaps":{"enableWidget":true}} + toolConfig.retrievalConfig.latLng → groundingChunks[].maps {placeId, uri, reviewSnippets}; $14/1k queries (3.x) |
no |
| Fetch a URL | — | {"type":"web_fetch_20260318","name":"web_fetch","max_uses":3,"citations":{"enabled":true}} (URLs already in context) |
— (web search open_page) |
{"urlContext":{}} (≤20 public URLs named in the prompt) → urlContextMetadata.urlMetadata[]; free tool, content = input tokens |
no |
| Run Python | {"type":"code_interpreter","container":{"type":"auto","memory_limit":"1g"}} → code_interpreter_call {code, outputs}; container files; per-session billing |
{"type":"code_execution_20260521","name":"code_execution"} → server_tool_use + bash_code_execution_tool_result {stdout, stderr, return_code, content[{file_id}]}; container reuse; container-hours |
{"type":"code_interpreter"} → code_interpreter_call {code, outputs[{type: logs, logs: "<JSON stdout/stderr/exit_code>"}]} (needs include: ["code_interpreter_call.outputs"]); no container object; $5/1k calls |
{"codeExecution":{}} → parts executableCode {language: PYTHON, code} + codeExecutionResult {outcome, output} (+ PNG inlineData); 30 s; no fee |
concept yes |
| Search your documents | {"type":"file_search","vector_store_ids":["vs_…"],"max_num_results":5} (Vector Stores API) |
pass search_result / document blocks with citations.enabled (your retrieval) |
{"type":"file_search","vector_store_ids":["<collection_id>"],"max_num_results":5} (Collections API; collections_search alias) → file_search_call {queries, results[{file_id, filename, score, text}]}; $2.50/1k + $0.10/GiB/day |
{"fileSearch":{"fileSearchStoreNames":["fileSearchStores/my-store"],"metadataFilter":"author = \"X\""}} (File Search stores; indexing $0.15/1M once) → groundingChunks[].retrievedContext |
OpenAI ↔ xAI literal; Gemini concept |
| Generate an image | {"type":"image_generation","quality":"low","size":"1024x1024"} → image_generation_call |
— | {"type":"image_generation"} → image_generation_call at Imagine per-image rates |
not a tool: generationConfig.responseModalities: ["TEXT","IMAGE"] on an image model |
OpenAI ↔ xAI |
| Call an MCP server | {"type":"mcp","server_label":"deepwiki","server_url":"https://mcp.deepwiki.com/mcp","require_approval":"never","allowed_tools":["read_wiki_structure"]} → mcp_list_tools, mcp_call, mcp_approval_request/response |
header anthropic-beta: mcp-client-2025-11-20; mcp_servers:[{type:url,url,name}] + tools:[{type:mcp_toolset, mcp_server_name, configs}] → mcp_tool_use/mcp_tool_result; no approvals |
{"type":"mcp","server_label":"deepwiki","server_url":"https://mcp.deepwiki.com/mcp","allowed_tools":["read_wiki_structure"]} → mcp_call {name, server_label, arguments, output, error}; no approvals (require_approval ignored); tokens only |
Interactions {"type":"mcp_server","name":"deepwiki","url":"https://mcp.deepwiki.com/mcp","allowed_tools":[…]} (documented; server names without -); generateContent mcpServers[] UNVERIFIED; SDK mcpToTool() runs calls client-side |
OpenAI ↔ xAI literal |
| Find a tool among many | {"type":"tool_search"} + defer_loading:true → tool_search_call/tool_search_output |
{"type":"tool_search_tool_regex_20251119","name":"tool_search_tool_regex"} + defer_loading → tool_search_tool_result |
{"type":"tool_search"} + defer_loading:true → 403 alpha users only |
— | concept |
| Let the model script tool calls | {"type":"programmatic_tool_calling"} + allowed_callers:["programmatic"] → program/program_output (JavaScript) |
code_execution_20260120+ + allowed_callers:["code_execution_20260120"] (Python; reply with tool_result + container) |
— | — (compositional function calling across turns) | OpenAI ↔ Anthropic concept |
| Shell on your machine | {"type":"shell","environment":{"type":"local"}} → shell_call {action.commands[]} → shell_call_output {output[{stdout,stderr,outcome}]} |
{"type":"bash_20250124","name":"bash"} → tool_use {command|restart} → tool_result |
{"type":"shell","environment":{"type":"local","skills":[…]}} → shell_call → shell_call_output {stdout, stderr, outcome} |
— (Antigravity agent sandbox only) | OpenAI ↔ xAI literal |
| Edit files | {"type":"apply_patch"} → apply_patch_call {operation, diff} → apply_patch_call_output |
{"type":"text_editor_20250728","name":"str_replace_based_edit_tool"} → tool_use {command: view|str_replace|create|insert} |
— | — (Antigravity built-ins write_to_file, replace_file_content, view_file…) |
concept |
| Drive a computer | {"type":"computer"} → computer_call {action|actions[], pending_safety_checks} → computer_call_output {computer_screenshot, acknowledged_safety_checks} |
{"type":"computer_toolset_20260801"} → member tool_use {name: screenshot|click|type|zoom…} → tool_result {toolset_name: "computer", content:[image]} |
— | {"computerUse":{"environment":"ENVIRONMENT_BROWSER","enablePromptInjectionDetection":true}} → functionCall {name: click, args:{x,y (0–999), intent}} (+ safety_decision) → functionResponse with screenshot inlineData (+ safety_acknowledgement) |
concept |
| Persistent notes | — (Agents API sessions) | {"type":"memory_20250818","name":"memory"} |
— | — | no |
| Ask a stronger model | — | {"type":"advisor_20260301","name":"advisor","model":"claude-opus-5"} (beta) |
— | — | no |
6. Server-tool loop semantics that differ
| Topic | OpenAI | Anthropic | xAI | Gemini |
|---|---|---|---|---|
| Where hosted results land | items in the same output[]; one call completes the request (max_tool_calls caps built-in calls) |
*_tool_result blocks in the same content[]; ~10 iterations → stop_reason: pause_turn → resend as-is |
items in the same output[]; the server loops inside one request up to max_turns (default = server cap; docs 1–2 quick, 3–5 balanced, 10+ deep); a client-side function/shell call ends the request and the follow-up gets a fresh budget |
parts in the same candidates[].content (executableCode/codeExecutionResult, groundingMetadata, urlContextMetadata); Gemini 3 exposes built-in calls as toolCall/toolResponse parts when toolConfig.includeServerSideToolInvocations: true (echo them back) |
| Mixed server + client calls in one turn | hosted calls never batched with function calls | possible (stop_reason: tool_use with an unanswered server_tool_use) |
possible: server tools run first, a client call ends the turn | Gemini 3 only ("tool combination", Preview; mode VALIDATED); Live API allows googleSearch + functions only |
| Approvals | MCP require_approval (default always) → mcp_approval_request item |
none in Messages (Managed Agents permission_policy) |
none (require_approval silently accepted) |
computer use safety_decision: require_confirmation → safety_acknowledgement: true; no MCP approvals |
| Result replay | keep hosted items in context (auto with previous_response_id) |
replay encrypted_content/encrypted_index unchanged (400 otherwise) |
keep items / use previous_response_id; replay reasoning.encrypted_content for cache hits |
echo executableCode/codeExecutionResult parts with their id and thoughtSignature; groundingChunks accumulate across stream chunks |
| Errors | mcp_call.error {type}; code_interpreter_call.status: failed |
*_tool_result_error {error_code} inside a 200 |
file_search_call.status: failed (collection not indexed → 404 on search); mcp_call.error string; tool type unknown → 422 bare string |
codeExecutionResult.outcome: OUTCOME_FAILED|OUTCOME_DEADLINE_EXCEEDED; urlRetrievalStatus: URL_RETRIEVAL_STATUS_ERROR|PAYWALL|UNSAFE; finishReason: MALFORMED_FUNCTION_CALL|UNEXPECTED_TOOL_CALL|TOO_MANY_TOOL_CALLS |
| Billing counters | tool_usage {web_search: {num_requests}} (undocumented); container sessions |
usage.server_tool_use {web_search_requests, web_fetch_requests}; container hours |
usage.server_side_tool_usage_details {web_search_calls, x_search_calls, code_interpreter_calls, file_search_calls, document_search_calls, mcp_calls, image_generation_calls}, num_server_side_tools_used, cost_in_usd_ticks; only successful executions billed |
usageMetadata.toolUsePromptTokenCount (+ per-modality details); Interactions usage.total_tool_use_tokens, grounding_tool_count[]; search queries counted against the 5,000/month quota |
7. Model coverage (from generated/compatibility/*matrix*.json and tools.json)
| Tool family | OpenAI models (current) | Anthropic models (current) | xAI models | Gemini models |
|---|---|---|---|---|
| function / custom tools | all text models (38 listed) | all 13 active Claude models | all 7 Grok text models (+ voice session, /v1/messages) |
21 listed (all 3.x / 2.5 text models, Live models, Gemma unknown) |
| web search | 43 models | all 13 | all 7 (Responses only) | 26 listed incl. gemini-3.1-flash-image (web + image search); Live models |
| code execution | 36 models | all 13 | all 7 | 17 (all 3.x / 2.5 text models; not Live) |
| MCP | 42 models | all 13 (beta header) | all 7 (+ voice) | mcpServers schema lists 135 (UNVERIFIED); SDK-side MCP 83 |
| file search / RAG | vector stores on Responses models | — | all 7 (file_search/collections_search) |
15 (fileSearch; 3.1 Pro 'AI Studio only') |
| tool search | 11 (gpt-5.4+, 5.5, 5.6, 6) | 12 | 7 listed, all 403 (alpha) | — |
| computer use | 11 | toolset GA 7; beta 11 | — | 9 (3.8/3.7/3.6/3.5 Flash, 3.5 Flash-Lite, 3 Flash Preview, 2.5 CU legacy, robotics) — PREVIEW |
| shell / editor / skills | 14–15 | all 13 (bash, editor); skills with code execution | shell (local) all 7; skills 404 | — (Antigravity only) |
| programmatic tool calling | not enumerated (docs: GPT-5.4+/6) | 12 (not Haiku 4.5) | — | — |
| image generation tool / modality | 31 | — | all 7 (image_generation) |
image models only (responseModalities) |
| URL context / web fetch | — | all 13 (web_fetch_*) |
— | 16 (urlContext) |
| Maps / X search | — | — | all 7 (x_search) |
15 (googleMaps, not Live) |
8. Portability checklist
- Keep tool definitions provider-neutral (name, description, JSON Schema with
additionalProperties: false, all fields required); mapparameters↔input_schema↔parameters/parametersJsonSchema; Gemini's OpenAPI-subsetparametersuses uppercasetype(OBJECT,STRING) — useparametersJsonSchemato keep one schema. - Normalise the call record to
{id, name, args(object)}— parseargumentson OpenAI/xAI; keep Gemini'sthoughtSignaturealongside the call. - Emit results as
function_call_outputitems (OpenAI/xAI), a leading user turn oftool_resultblocks (Anthropic), or one user Content offunctionResponseparts after all calls of the step (Gemini); represent errors as text (OpenAI/xAI),is_error: true(Anthropic) or an error object inresponse(Gemini). - Loop until no tool call: OpenAI/xAI → no
function_callitem inoutput[]; Anthropic →stop_reason != tool_use && != pause_turn; Gemini → nofunctionCallpart (checkfinishReasonforMALFORMED_FUNCTION_CALL/MISSING_THOUGHT_SIGNATURE). - Keep
toolsidentical across turns on all four (cache prefix; Gemini Interactions and xAIprevious_response_iddo not need them resent but accept changes); restrict withallowed_tools(OpenAI) /allowedFunctionNames(Gemini) rather than editing the list. - Bound the loop (
max_turnson xAI is server-side; use your own counter elsewhere) and honour approvals/safety checks — reference implementations:examples/shared/tool-loop/{openai,anthropic,gemini}_tool_loop.{py,ts},examples/xai/responses/tool_loop.py,examples/xai/tools/function-calling/loop.ts(allLIVE_VERIFIED).
Related: features · state-management · streaming · docs/openai/tool-loop.md · docs/tools/anthropic/tool-use-loop.md · docs/xai/tool-loop.md · docs/gemini/tool-loop.md · docs/tools/{openai,anthropic,xai,gemini}/index.md.