# Run a multi-step tool-calling loop

> Run a multi-step tool-calling loop: 9 companies fully do it, 25 with a limit, 42 carry the tag. Every listing in Model APIs & inference with the reason and a source.

- Canonical: https://www.anchorterminal.com/niches/inference.tool-loop/
- Markdown: https://www.anchorterminal.com/niches/inference.tool-loop/index.md (~3,450 tokens)
- Slim: https://www.anchorterminal.com/niches/inference.tool-loop/index.min.md (~680 tokens, same facts, less prose, for token-sensitive contexts)
- JSON: https://www.anchorterminal.com/niches/inference.tool-loop/index.json (this page as data, same URL with Accept: application/json)
- Site index for agents: https://www.anchorterminal.com/llms.txt (the key docs and every listing in one file, at most 16 MB, about 4 million tokens: https://www.anchorterminal.com/llms-full.txt)
- API: https://www.anchorterminal.com/api/v1/index.json
- MCP: https://www.anchorterminal.com/mcp (the whole catalogue as tools, no key)
- DNS: dig +short TXT <slug>.grade.anchorterminal.com (the grade as a DNS TXT record, for a sandbox with no HTTP where names still resolve; also top., best.<category>. and <a>.vs.<b>. under the zone, and the same under grade.letme.dev)
- Licence: CC BY 4.0, credit Anchor Terminal and link this page (https://www.anchorterminal.com/terms/#data)
- Updated: 2026-11-15

Four or more companies' docs show a full fit.

| Companies, full fit | With a limit | Carry the tag | Share of the category's work | First-hand accounts |
| --- | --- | --- | --- | --- |
| 9 | 25 | 42 | 24 | 12 |

## What counts as done

Every tool call comes back as a structured tool-call object with a parseable name and arguments, and the echoed assistant turn is accepted on the next request.

Measured, for our own tests: a 30-turn loop finishes without a 400, an empty tool_calls array or the same call repeated more than three times.

How agents ask for it: run this task with my tools until it is done; call the model, run the tool it asks for, send the result back, repeat; drive a coding agent or MCP tools from a model API

## Does the job

The docs show it meets the test above, and nothing we read says otherwise.

- **Amazon Bedrock (Amazon Web Services, Inc.)**: Converse returns toolUse blocks with toolUseId, name and input, stopReason tool_use, and toolResult goes back by id; unsupported toolChoice fails with a named ValidationException rather than silently. The Anthropic Messages route serves tool_use natively.
- **Claude API (Anthropic, PBC)**: tool_use blocks carry name and JSON input, tool_result goes back by id, and interleaved thinking is automatic; acc-24 ran a multi-agent security platform on it and acc-25 ran 20-iteration MCP loops.
- **Azure OpenAI in Foundry Models (Microsoft Corporation)**: Chat Completions returns tool_calls with id, name and arguments, parallel calls on supporting deployments, and the Responses API is available on the v1 endpoint; every response carries x-ratelimit headers and 429s carry retry-after-ms.
- **BlockRun.AI (BlockRun, Inc.)**: Chat Completions takes tools and tool_choice (auto, none, required or named), /v1/messages and /v1/responses are also served, usage lists completion tokens including reasoning, and failed calls (400, 402, 429, 5xx) are not charged.
- **Cerebras Inference (Cerebras Systems Inc.)**: tool_calls with strict constrained decoding, per-model columns for tool_choice, parallel and multi-turn calling, and tool_call_id round trips; the trial tier's request limits apply before the Developer tier.
- **DeepInfra (Deep Infra Inc.)**: tool_calls with single, parallel and streaming calls on supported models, 200 concurrent requests per model, and fail_fast to get an immediate 429 instead of queueing at capacity.
- **Meta Model API (Meta Platforms, Inc.)**: Chat Completions returns tool_calls and Responses returns function_call items with previous_response_id carrying reasoning across turns; 3,000 requests and 4M tokens a minute per model, 429s with Retry-After and x-ratelimit headers.
- **Mistral AI API (Mistral AI (RCS Paris 952 418 325))**: Chat Completions returns tool_calls with finish_reason tool_calls, tool_choice any forces a call, up to 128 tools per request, and the reasoning guide says to replay the full assistant message including ThinkChunk each turn.
- **OpenAI API (OpenAI OpCo, LLC)**: The Responses API returns function_call items with call_id and strict function schemas, reasoning items carry encrypted_content that must be passed back with tool outputs, and x-ratelimit headers plus Retry-After guide backoff.

## Does the job, with a limit

It does the job, but something in the docs or a user's account will get in an agent's way. The limit is named in each row.

- **Antseed (Antseed Foundation)**: The proxy serves Chat Completions, Responses and Messages and translates between them, and the guides target Claude Code and OpenCode, but tool-call shape rests on whichever independent seller answers and no tool fidelity is documented.
- **Cloudflare AI Gateway (Cloudflare, Inc.)**: Provider-native and OpenAI-compatible routes pass tool calls through unchanged, but dynamic routes accept the OpenAI chat shape only and Unified Billing traffic is capped at 200 requests per 60 seconds per gateway.
- **Cloudflare Workers AI (Cloudflare, Inc.)**: Models flagged for function calling return a tool_calls array, but text generation is capped at 300 requests a minute and frontier models that need Workers Paid at 20 (50 with prepaid gateway credits), which a 50-call turn can exceed.
- **Cohere Chat API (Command models) (Cohere Inc.)**: v2 chat returns tool_calls with ids, supports parallel and multi-step calls and documented tool_call_id echoes, but Command A+ and the newer variants stay at trial limits of 20 requests a minute and 1,000 calls a month even on a production key.
- **DeepSeek API (Hangzhou DeepSeek Artificial Intelligence Co., Ltd.)**: tool_calls with a beta strict mode and documented reasoning_content echo rules, but acc-09 saw deepseek-v4-pro repeat the same git status call for 50 rounds and 8.9M tokens with no provider-side stop and no cache hits.
- **Eden AI (Eden AI SAS)**: Chat completions take tools and tool_choice across a thousand models, but every account is capped at 10 requests a second shared across keys and reasoning models need max_tokens of 500 or more or return empty.
- **Glimpse (Elorian AI, Inc.)**: The single Together endpoint lists tools and tool_choice (none, auto, required but not a named function) and 262K context, but Glimpse is a visual reasoning research preview with no first-hand account of a tool loop.
- **Fireworks AI Inference (Fireworks.ai, Inc.)**: tool_calls with streaming argument deltas and parallel calls on supporting models, but the rate-limit page says traffic within limits can still be load shed with 503 Service Overloaded, which a 30-turn loop must absorb.
- **Gemini Developer API (Google LLC)**: Function calls arrive as function_call parts or steps, but thought signatures must be kept and resent in stateless mode; acc-08 saw 400 after the first tool call when a proxy dropped them, and acc-16 saw free-tier sessions stuck on 429.
- **GroqCloud (Groq LLC)**: tool_calls with ids that must match on the way back and a per-model support table, but the free plan allows 8K tokens a minute and 1,000 requests a day on gpt-oss-120b, so a 100K-token loop needs the Developer plan.
- **Hugging Face Inference Providers (Hugging Face, Inc.)**: Tools and tool_choice pass through to the chosen provider, but supports_tools is true on 174 of 247 provider endpoints in the router list, so the guide tells callers to pin a provider with model:provider to cut variance.
- **Jais 3 (Mohamed bin Zayed University of Artificial Intelligence (Institute of Foundation Models), with Inception42 (a G42 company))**: vLLM with --enable-auto-tool-choice and --tool-call-parser k2_horizon returns structured tool calls and the card ships a loop example, but every earlier turn must carry reasoning_content or the template errors, and acc-13 and acc-19 show vLLM parsers failing on other reasoning models.
- **LiteLLM (Berrie AI Incorporated)**: Tool calls pass through in OpenAI shape, but acc-07 got a Qwen3 tool call as raw JSON in content through the proxy (closed not planned) and acc-08 got 400 after the first Gemini tool call when the proxy dropped thought_signature.
- **MiniMax API (Nanonoble Pte. Ltd.)**: Both formats return structured tool_calls or tool_use blocks and the docs say to echo the full assistant content with thinking blocks; tool_choice is auto only, stop_sequences are ignored, and acc-11 saw OpenAI-shaped tool calls reach Claude Code through a gateway.
- **Muse Spark Open (Meta Platforms, Inc.)**: Tool calls come back as standard tool_calls only when the server runs the muse reasoning and tool-call parsers (vLLM also needs --enable-auto-tool-choice); reasoning can't be turned off and defaults to xhigh.
- **Nemotron 3.5 Super VL (NVIDIA Corporation)**: The API reference lists tools, tool_choice, parallel_tool_calls and a separate reasoning_content; acc-20 saw a tool call land in the body with empty tool_calls on the same hosted catalogue, and the trial endpoint is for prototyping only.
- **Novita AI (Novita AI)**: Chat Completions returns tool_calls, puts reasoning in reasoning_content, carries usage in the last stream chunk and honours stop; tool support is per model and the guide's example model is old, so check the model page first.
- **OpenRouter (OpenRouter, Inc.)**: Tool calls arrive in OpenAI shape across providers, but routing to tool-capable hosts is best effort unless require_parameters is set; acc-04 saw hosts return function-call markup as text with a 200 and acc-02 a generic 'Provider returned error'.
- **Prism Inference (Prism Technologies Inc)**: Chat Completions, Responses and Messages carry function tools or tool_use blocks with tool_choice, include_usage and reasoning controls; Responses is stateless with no previous_response_id, and only two models are served.
- **Router.com (Ramp Router) (Ramp Business Corporation)**: Responses and Messages carry function and custom tools, usage.cost and the serving provider in the body and X-Router-Provider; capability translation means the least capable model in a fallback list bounds tools and context.
- **Beam 1.0 (Reflection AI)**: tool_calls with finish_reason tool_calls, tool_choice required and parallel calls; every assistant message must be returned with its reasoning_content unchanged or the request fails with 400 missing_required_parameter, and effort defaults to max.
- **RouterPlus (EVO HQ Inc)**: Chat Completions, Responses and Messages pass tool_calls and tool_use; max_tokens must be sent or the balance hold assumes 32,768 output tokens and can return insufficient_quota, and there is no idempotency key so a retry bills twice.
- **SambaCloud (SambaNova Systems, Inc.)**: tool_calls with forced tool_choice on Chat Completions, function_call items on a stateless Responses API, and tool_use on Messages; the free tier allows 20 requests a day per model, so a loop needs the Developer tier.
- **SiliconFlow (SiliconFlow Labs Pte. Ltd.)**: Chat Completions returns tool_calls and a separate reasoning_content with thinking_budget on Qwen3; the function-calling guide's model list is stale (DeepSeek-V2.5) and tool support is per model.
- **Together AI Inference (Together Computer, Inc.)**: tool_calls with parallel calls on supported models, a reasoning_content the example returns faithfully, and a Messages API with tool_use; tool support is per model and serverless 429s and 503s are best-effort capacity.
- **Vercel AI Gateway (Vercel Inc.)**: Chat Completions, Responses and Anthropic Messages pass tools and tool_choice, and providerMetadata names the resolved provider; default routing can move a conversation between providers, which the sessionAffinity option exists to limit.
- **Vertex AI (Gemini Enterprise Agent Platform) (Google LLC)**: Gemini thinking models return thought signatures that must be preserved across multi-turn and multi-step function calls; the function-calling page itself wasn't found at the paths tried, so tool-call shape rests on the thinking guide.
- **xAI API (Grok) (SpaceXSI (xAI))**: The Responses API returns function_call items and encrypted reasoning to pass back; Chat Completions is deprecated and has no field for encrypted reasoning, so multi-turn tool loops there lose state; per-second limits are RPM divided by 48.

## Couldn't check

The docs were closed to us, or silent. These carry no mark either way.

- **StepFun API (STEPFUN TECHNOLOGY (HK) LIMITED)**: The host platform.stepfun.ai closed the connection on every attempt (curl exit 56, no HTTP status) from this session, so no page was read.
- **Tencent Cloud TokenHub (Tencent Cloud Computing (Beijing) Co., Ltd. (腾讯云计算（北京）有限责任公司))**: The host cloud.tencent.com closed the connection on every attempt (curl exit 56, no HTTP status) from this session, so no page was read.

- not this job: Simplismart (Verute Technologies Private Limited): The Llama 3.1 8B OpenAPI file has no tools or tool_choice parameter and the quickstart shows text completion only.
- not this job: SWE-2 Flash (Cognition AI, Inc.): No model API; SWE-2 Flash runs only inside Devin Desktop and Devin CLI, so there is no tool-call object for a harness to parse.

6 listings in the category were added after this reading and aren't marked yet.

## What stops agents finishing

- tool calls returned as text or inside a reasoning block, so the harness never executes them
- provider rejects the echoed tool-call turn (missing thought_signature or reasoning_content) so every later request fails
- model repeats the same tool call indefinitely with no provider-side stop and no cache hits, running up 8.9M tokens

## Around it

Before: search.find-url, web-scrapers.page-to-markdown. After: agents.sandbox-run, obs.trace.

## Where this comes from

Weight 24 of 100 in Model APIs & inference: 12 of 28 accounts describe a tool loop failing or succeeding on a model API; the most common job in the issues read

As data: https://www.anchorterminal.com/api/v1/use-cases/inference.tool-loop.json, or the MCP tool get_use_case with id inference.tool-loop. Draft, read 2026-11-12 from public docs and first-hand accounts, not yet reviewed by a person.

---

Query this catalogue directly over MCP: https://www.anchorterminal.com/mcp (Streamable HTTP, no key). How: https://www.anchorterminal.com/agents/#mcp
