Niches · Model APIs & inference

Put a different model behind an existing agent harness

Four or more companies' docs show a full fit.

7companies, full fit
28companies, fit with a limit
60companies carry the tag
6share of the category's work
4first-hand accounts

What counts as done

Tool calls arrive in the harness's own format, usage comes in the final stream chunk, and 429s are returned as 429s so the harness backs off.

Measured, for our own tests: the harness completes a 30-file edit or a 50-tool-call session. No docs we read state these figures, so they don't count against a listing here.

How agents ask for it: run Claude Code against DeepSeek or MiniMax, point OpenCode at my vLLM server, use the Anthropic-compatible endpoint so I don't rewrite the client

Does the job 7

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

ListingWhy
Amazon Bedrock A
Amazon Web Services, Inc.
Anthropic Messages at /anthropic/v1/messages on bedrock-runtime and bedrock-mantle with an API key, plus OpenAI Chat Completions and Responses; bedrock-runtime has no GET /models and mantle path prefixes vary by model. source
Azure OpenAI in Foundry Models A
Microsoft Corporation
OpenAI v1 endpoint at <resource>.openai.azure.com/openai/v1 takes the OpenAI SDK with the deployment name as model, serves Chat Completions and Responses, and returns 429 with retry-after-ms and x-ratelimit headers. source
Cloudflare AI Gateway B
Cloudflare, Inc.
Provider-native routes include /anthropic/v1/messages (provider key in x-api-key) and an OpenAI-compatible endpoint with {provider}/{model} ids, so Claude Code or an OpenAI harness points at the gateway unchanged. source
DeepInfra B
Deep Infra Inc.
Anthropic Messages at /anthropic/v1/messages with Claude Code model mapping and streaming, OpenAI Chat Completions and Responses, and the final stream chunk carries usage. source
DeepSeek API D
Hangzhou DeepSeek Artificial Intelligence Co., Ltd.
Anthropic format at api.deepseek.com/anthropic with a Claude Code guide, stop_sequences and stream fully supported, claude-* model names mapped to V4 Pro or Flash, plus OpenAI Chat Completions with stream_options and Responses. source
Fireworks AI Inference BB
Fireworks.ai, Inc.
Anthropic /v1/messages compatibility with usage in streaming responses, tool-search translation and documented unsupported fields, plus OpenAI Chat Completions and Responses; 429s carry Retry-After. source
Prism Inference C
Prism Technologies Inc
All three wire formats with streaming, tools and stop in the OpenAPI, 429 returned as 429 with Retry-After and a retryable flag, include_usage supported, and agent guides for Codex, OpenCode, Hermes and OpenClaw. source

Does the job, with a limit 30

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.

ListingWhy
Claude API A
Anthropic, PBC
An OpenAI SDK compatibility layer serves Chat Completions with tools, stream_options and stop sequences, but strict is ignored, prompt caching is unavailable there, and forced tool_choice returns 400 on Opus 5.5, Sonnet 5.5 and Fable 5.x. source
Antseed C
Antseed Foundation
Local proxy at localhost:8377 speaks OpenAI Chat Completions, Responses and Anthropic Messages and translates formats for the seller, with no API key needed; usage in the final chunk and 429 behaviour are not documented. source
BlockRun.AI A
BlockRun, Inc.
Serves Chat Completions, Anthropic Messages and Responses with streaming and usage, and ships an MCP server for Claude Code, Cursor and Codex, but a harness pointing straight at the gateway must sign x402 payments on 402 or hold an api.blockrun.ai key. source
Cerebras Inference B
Cerebras Systems Inc.
OpenAI Chat Completions at api.cerebras.ai/v1 with tools and a documented unsupported-parameter list, but no Anthropic Messages endpoint, and the 429 names the token bucket without documented rate-limit headers. source
Cloudflare Workers AI BB
Cloudflare, Inc.
OpenAI Chat Completions at /ai/v1/chat/completions, Responses only for the two gpt-oss models, and no Anthropic Messages; 429 bodies carry code 3036 for the free allocation and 3040 for capacity. source
Cohere Chat API (Command models) B
Cohere Inc.
A Compatibility API at api.cohere.ai/compatibility/v1 serves Chat Completions with tools, strict and structured outputs and streaming, but there is no Anthropic Messages endpoint and usage in the final chunk is not shown. source
Eden AI C
Eden AI SAS
OpenAI-compatible chat completions only in the pages read; a Claude Code integration page exists but was not opened, and 429 behaviour is not documented. source
Glimpse F
Elorian AI, Inc.
Reached through OpenRouter's Chat Completions and Responses APIs; no Anthropic Messages endpoint was found, and errors after a 200 arrive inside the body or SSE stream rather than as status codes. source
Gemini Developer API B
Google LLC
An OpenAI-compatible endpoint serves function calling and maps reasoning_effort to thinking_level, but there is no Anthropic Messages endpoint and acc-06 reported fenced JSON on the compatible endpoint. source
GroqCloud BB
Groq LLC
OpenAI Chat Completions and Responses at api.groq.com/openai/v1 with retry-after and x-ratelimit headers on every call, but no Anthropic Messages endpoint and free-plan token limits below an agent loop. source
Hugging Face Inference Providers C
Hugging Face, Inc.
OpenAI Chat Completions at router.huggingface.co/v1 with stream_options include_usage, and a beta Responses API, but no Anthropic Messages endpoint and no documented 429. source
Jais 3 C
Mohamed bin Zayed University of Artificial Intelligence (Institute of Foundation Models), with Inception42 (a G42 company)
Served through vLLM or SGLang's OpenAI endpoint with the k2_horizon reasoning and tool parsers; no Anthropic format, and --trust-remote-code runs repository Python on the host. source
LiteLLM B
Berrie AI Incorporated
/v1/messages serves the Anthropic format over every provider with streaming, fallbacks and a Claude Code setup page, but acc-07 and acc-08 show tool calls degraded in transit for Ollama and Gemini backends. source
Meta Model API D
Meta Platforms, Inc.
Chat Completions, Responses and Anthropic Messages are all served, but the Messages route returns 400 for stop_sequences, top_k, a named tool_choice and thinking disabled, which an Anthropic-format harness may send. source
MiniMax API C
Nanonoble Pte. Ltd.
Anthropic Messages, Chat Completions and Responses are served and usage arrives in the last stream chunk, but stop_sequences and top_k are ignored, tool_choice is auto only, and acc-11 reports mis-shaped tool calls and 503s via a gateway. source
Mistral AI API B
Mistral AI (RCS Paris 952 418 325)
Chat Completions and a stateless Responses API, no Anthropic Messages; stream_options.include_usage must be set or usage is absent from stream events; 429 carries Retry-After. source
Muse Spark Open C
Meta Platforms, Inc.
OpenAI Chat Completions through vLLM, SGLang or TensorRT-LLM with the muse parsers; no Anthropic Messages endpoint, and stopping on <|eom|> truncates tool calls. source
Nemotron 3.5 Super VL C
NVIDIA Corporation
OpenAI Chat Completions only, with stream_options.include_usage and stop; self-hosting needs --reasoning-parser nemotron_3 and --tool-call-parser qwen3_coder; acc-20 shows tool calls arriving as text on the hosted route. source
Novita AI C
Novita AI
Anthropic-format base URL api.novita.ai/anthropic with a Claude Code guide, plus OpenAI Chat Completions with include_usage in the last chunk; no Retry-After header and no 429 body beyond the error code. source
OpenAI API AA
OpenAI OpCo, LLC
Native Chat Completions and Responses, no Anthropic Messages; 429 with Retry-After and x-ratelimit headers, but the newest models call tools only on Responses, so a Chat Completions harness loses them. source
OpenRouter B
OpenRouter, Inc.
OpenAI Chat Completions and Responses with usage and provider in the stream and Retry-After on errors, but mid-stream failures come as error events inside a 200; the /messages endpoint from the listing's note wasn't re-read. source
Router.com (Ramp Router) C
Ramp Business Corporation
Responses, Chat Completions and Anthropic Messages, but Chat Completions lacks fallback lists and usage.cost and needs include_usage; 429 says respect retry-after; errors use the OpenAI envelope or Anthropic format. source
Beam 1.0 C
Reflection AI
Chat Completions and Models only, no Responses or Messages; stop is accepted and ignored; include_usage supported; 429 carries Retry-After and x-should-retry false when the daily allowance is spent. source
RouterPlus D
EVO HQ Inc
OpenAI Chat Completions, Responses and Anthropic Messages, with a typed 429 and retry-after in both the OpenAI code and the Anthropic rate_limit_error shape; unknown parameters are dropped and named in x-tm-dropped-params. source
SambaCloud B
SambaNova Systems, Inc.
Chat Completions, Responses and Anthropic /v1/messages with count_tokens; headers report RPM and RPD; gpt-oss-120b needs the Chat Completions tool_choice shape even on Responses. source
SiliconFlow E
SiliconFlow Labs Pte. Ltd.
OpenAI Chat Completions plus an Anthropic /v1/messages reference with tools, stop_sequences and stop reasons, and a Claude Code guide; 429 names RPM, RPD, TPM or TPD but the rate-limits page is 404. source
Simplismart F
Verute Technologies Private Limited
OpenAI-compatible base URL with stream, stop and max_tokens; the OpenAPI file has no usage field, 429 format or rate limits, so a harness may hang on the final chunk. source
Together AI Inference B
Together Computer, Inc.
Chat Completions, Responses and Anthropic Messages (base URL api.together.ai) with tool_use, thinking and streaming; 429 and 503 both signal demand so a harness must back off on both; usage in the final chunk not verified. source
Vercel AI Gateway B
Vercel Inc.
OpenAI Chat Completions, Responses and Anthropic Messages (with cache_control, thinking, stop_sequences) by base URL change; 429 with retry-after as seconds or an HTTP date; usage in the final chunk not verified. source
xAI API (Grok) BB
SpaceXSI (xAI)
Responses is the supported path; Chat Completions is deprecated and the Anthropic SDK compatibility fully deprecated; 429 is returned as 429 under RPS and TPM caps but no Retry-After is documented. source

Couldn't check 3

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

ListingWhy
StepFun API E
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 C
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.
Vertex AI (Gemini Enterprise Agent Platform) AA
Google LLC
The OpenAI-compatible and Anthropic raw endpoints weren't opened (the OpenAI library page returned 404 at the paths tried).

Doesn't do this job 1

Listings in the category that do something else, with the reason.

ListingWhy
SWE-2 Flash F
Cognition AI, Inc.
Only Devin's own harness; the short name swe resolves to the latest family model, not necessarily Flash. source

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

What stops agents finishing

  • tool calls arrive in OpenAI shape on an Anthropic-format endpoint and the harness fails silently
  • final stream chunk omits usage so the harness hangs
  • generic 503 instead of 429 so the harness retries hard and spends the quota
  • tool parser crashes on a streamed call (IndexError in the serving stack)

Around it

After: inference.tool-loop.

Where this comes from

Weight 6 of 100 in Model APIs & inference: 4 of 28 accounts, all from people wiring Claude Code or another harness to a non-native model; two are vendor-adjacent

The same as data: /api/v1/use-cases/inference.harness-compat.json, or the MCP tool get_use_case with id inference.harness-compat. Other jobs in this category: /api/v1/niches?category=inference.

Draft, read 2026-11-12 from public docs and first-hand accounts, not yet reviewed by a person. A mark says what the docs show, not what a run measured.

For companies

Do agents find, use and choose your tools?

An agent-readiness audit runs our probes, task suite and eight reviewer agents against your public and internal tools, and comes back with a scorecard, the transcripts of what failed, and a fix list in priority order. From $2,500, re-run included. We never take payment to move a rank. We do help companies earn one.