GuidesRunning & maintaining

OpenRouter: No endpoints found that support tool use

Why OpenRouter says no endpoints support tool use or image input, or that reasoning is mandatory, and how to check a model before an agent uses it.

November 14, 2026The Everpod team
The short answer

OpenRouter sends each request to one of the providers that host the model, and these messages mean none of them can take this request. “No endpoints found that support tool use” (404): the agent sent tools, as agents always do, and no provider of that model accepts them. “No endpoints found that support image input” (404): an image reached a text-only model. “Reasoning is mandatory for this endpoint and cannot be disabled” (400): thinking was switched off for a model that always thinks. “No allowed providers are available for the selected model” (404): your provider restrictions exclude every host. Check the model’s supported_parameters, input types and reasoning block in OpenRouter’s models API, then change the model, the session or the setting. Retrying changes nothing.

No endpoints found that support tool use

No endpoints found that support tool use. To learn more about provider
routing, visit: https://openrouter.ai/docs/provider-routing
Copied

A coding or personal agent works by calling tools (reading files, running commands, searching), so every request it makes carries their definitions. Not every provider that hosts a model accepts them. OpenRouter’s provider routing makes “a best effort to route to providers known to support tool use”; when the model has no such provider, the request ends here. Free variants and small or specialised models are where agents meet it: Cline users got it on nvidia/nemotron-3.5-lightning in August 2026, and Claude Code Router users on qwen/qwen3-coder in 2025, when one of its providers took no tools.

The same model can work in one tool and fail in another. In that Claude Code Router thread the maintainer explains that Cline then described its tools inside the prompt rather than sending them as the API’s tools parameter, so the provider never saw a tools request. An agent that uses the native parameter cannot do without it, so the fix is the model: pick one with at least one endpoint that lists tools (checked below).

Watch for it in scheduled jobs. In a Hermes report from June 2026, a cron job set to nousresearch/hermes-4-70b hit this 404, fell back to openrouter/auto, and the run was recorded as successful under the original model name. If a job’s output changes character without an error, look at which model actually answered.

Checking a model before you give it to an agent

OpenRouter’s models API describes each model with the fields these messages turn on: supported_parameters, the parameters that work with it (tools among them), and architecture.input_modalities, the input types it takes (text, image, file). One model, and one model provider by provider:

curl -s https://openrouter.ai/api/v1/model/qwen/qwen3-coder | jq
curl -s https://openrouter.ai/api/v1/models/qwen/qwen3-coder/endpoints | jq
Copied

Neither needs a key. The second lists each endpoint with its provider_name and its own supported_parameters, which is what matters for the tool-use message: on November 14, 2026 qwen3-coder had two endpoints, DeepInfra and Venice, both with tools, and its input types were ["text"] only, so a screenshot sent to it would fail. To list only models that take tools, filter the list with ?supported_parameters=tools, on the API or on the models page. Provider lists change, free variants most of all, so check the model you are about to use rather than one you checked last month.

No endpoints found that support image input

An image went to a model whose input types are text only. Agents send images more often than their owners expect: a screenshot pasted into the chat, a browser tool’s capture, an image file the agent read. OpenClaw users saw it as “HTTP 404: No endpoints found that support image input” on a non-vision OpenRouter model.

The part that confuses people comes after. In some agents the refused image stays in the conversation, and since every request resends the history, every later message fails too, including plain text. A jcode report puts it as “every subsequent turn of the session is poisoned”, and Claude Code pointed at a text-only model through its base URL shows the same thing with that model’s own refusal: later prompts fail, and so does compaction, because the compaction request carries the image as well. The way out:

A different refusal, an image that is too large or in a format the provider rejects, arrives instead as “Provider returned error” with the provider’s own reason inside it.

Reasoning is mandatory for this endpoint and cannot be disabled

Some models always reason, and refuse a request that asks them not to. OpenRouter’s reasoning documentation marks them with reasoning.mandatory in the models API: “When true, hide disable controls and do not send effort: "none" — the model rejects it.” Google’s Gemini 3.5 Flash is one, as read on November 14, 2026:

"reasoning": {
  "mandatory": true,
  "default_enabled": true,
  "supported_efforts": ["high", "medium", "low", "minimal"],
  "default_effort": "medium"
}
Copied

The request that trips it carries reasoning: {effort: "none"} or reasoning: {enabled: false}, and agents send that in two ways. One is a thinking level you set to off. The other is the agent’s own side calls: in Claude Code Router, background tasks and subagents go out with thinking disabled, which its OpenRouter transformer turned into enabled: false, so stepfun/step-3.5-flash:free failed on those calls while ordinary turns worked. Either way the fix is to stop sending off: choose a low level instead (the model’s supported_efforts lists what it takes), or use a build of the tool that leaves the field out when thinking is disabled. Asking for the reasoning to be hidden is not the same thing: with exclude: true the model still reasons and the tokens are still billed.

In OpenClaw, set a level the model accepts with /think low in the conversation, or in agents.defaults.thinkingDefault for every new one (the thinking-level guide has the order in which those settings win). OpenClaw also sent off without being asked: on 2026.10.10 and 2026.11.2, right after a Gateway restart, its OpenRouter model entries could say the model does not reason, every level was clamped to off, and the request went out with effort: "none". The report narrows the trigger to OpenRouter behind a custom base URL, such as a local proxy, and the log shows “Thinking level clamped to off for openrouter/<model>” when it happens. A retry once the Gateway had loaded the model’s capabilities went through. The fix was merged on November 11, 2026, the day after 2026.11.3 was published, so expect it in the release after that.

No allowed providers are available for the selected model

{"error":{"message":"No allowed providers are available for the selected model.","code":404}}
Copied

Here the model has providers, and your own restrictions rule them all out. OpenRouter lets an account allow or ignore providers for every request, in its settings at openrouter.ai/settings/privacy, and a request can name its own with only or ignore; the provider selection docs say that when the account’s allowed list and a request’s only list share no provider, “the request fails with a 404.” In a Claude Code Router thread, enabling the needed provider in the OpenRouter account’s settings cleared it for qwen/qwen3-coder, and for a :free model, dropping the suffix did, at the paid model’s price. Compare the providers in the model’s endpoints list with the ones you allow, and widen the list or pick a model your list covers. A refusal that names your data policy or training settings instead comes from OpenRouter’s privacy controls, which can leave a model with no endpoint in the same way.

Give your AI agents a permanent cloud computer.

We set it up and secure it, so all you do is use them. Run Claude Code, Codex or any agent you use, more at once than your laptop can hold, and keep them working when it’s closed. Or have OpenClaw set up and run for you.

See the developer pod

From $24/mo · your own plan or API key

Get your OpenClaw agent

$29/mo, cancel anytime · model usage included