Tools and plugins

Choose who runs the tool

• Server tools — MiniRouter gets the current time, searches the web or reads a public web page, then returns the model's answer in one request.
• Your own functions — your app runs code, such as a database query, and sends the result back to the model.
• Plugins — MiniRouter shortens a long prompt or repairs JSON. The model does not call a plugin.

The examples use Chat Completions (POST /v1/chat/completions). MiniRouter's server tools and plugins are available on this endpoint only. Tools require a model that supports function calling; check tools in its supported_parameters in the catalog.

Try a server tool

Create an API key, save its recovery link, and add credits. Replace the placeholder below and run it in your terminal before running the examples on this page.

export MINIROUTER_KEY='paste-your-api-key-here'

Examples use paid models. For free requests, follow the Auto Free guide.

Run this in the same terminal. It asks Qwen 3.5 Flash to get the current year with minirouter:datetime.

curl https://api.minirouter.sh/v1/chat/completions \
  -H "Authorization: Bearer $MINIROUTER_KEY" \
  -H "Content-Type: application/json" \
  -d '{
  "model": "alibaba/qwen3.5-flash",
  "messages": [
    {
      "role": "user",
      "content": "Use the datetime tool to find the current year, then reply with only that four-digit year."
    }
  ],
  "max_tokens": 128,
  "reasoning": {
    "effort": "none"
  },
  "tools": [
    {
      "type": "minirouter:datetime"
    }
  ],
  "max_tool_calls": 2
}'

Read the year in choices[0].message.content. You do not need to write a datetime function or send a tool result yourself.

How the tool loop works

1. You send a prompt and a tool. The tools array tells the model what it can use.
2. The model requests a tool call. MiniRouter runs the built-in tool and sends its result back to the model.
3. The model answers. MiniRouter returns that answer in the same HTTP response. It can repeat step 2 within your max_tool_calls limit.

Listing a tool makes it available; the model may answer without using it. Ask explicitly when the task needs a tool, as in the example above. In a non-streaming response, x-minirouter-server-tool-rounds lists the model request IDs. Two or more rounds show that the server tool loop continued.

You can mix built-in tools with your own function definitions. When the model requests one of your functions, MiniRouter returns that call for your app to handle using the function-calling flow.

Available server tools

To set the timezone, replace the example's tools field with:

{
  "tools": [{
    "type": "minirouter:datetime",
    "parameters": { "timezone": "Europe/London" }
  }]
}

Read a web page. This request allows one fetch from example.com:

curl https://api.minirouter.sh/v1/chat/completions \
  -H "Authorization: Bearer $MINIROUTER_KEY" \
  -H "Content-Type: application/json" \
  -d '{
  "model": "alibaba/qwen3.5-flash",
  "messages": [
    {
      "role": "user",
      "content": "Use web_fetch to read https://example.com and summarize the page in one sentence."
    }
  ],
  "max_tokens": 256,
  "reasoning": {
    "effort": "none"
  },
  "tools": [
    {
      "type": "minirouter:web_fetch",
      "parameters": {
        "allowed_domains": [
          "example.com"
        ],
        "max_uses": 1,
        "max_content_tokens": 2000
      }
    }
  ],
  "max_tool_calls": 2
}'

Each successful fetch adds a $0.001 platform fee, plus model tokens. Failed, blocked and timed-out fetches have no fetch fee. Redirects count as one fetch. usage.server_tool_use reports the successful, billed count in web_fetch_requests. Datetime has no tool fee.

Web fetch requires a supported paid route with authoritative billing. Free models, provider keys (BYOK), and routes without authoritative billing reject it.

Web fetch reads a URL; it does not search the web or sign in to a website. Private-network addresses are blocked. The page text is sent to the model as tool output, so treat it as untrusted input.

Search the web

Search Formula 1's site for results published in the past month:

curl https://api.minirouter.sh/v1/chat/completions \
  -H "Authorization: Bearer $MINIROUTER_KEY" \
  -H "Content-Type: application/json" \
  -d '{
  "model": "zai/glm-5.3-flash",
  "messages": [
    {
      "role": "user",
      "content": "Search the web for the latest Formula 1 Grand Prix result. Give the winner and source URL."
    }
  ],
  "tools": [
    {
      "type": "minirouter:web_search",
      "parameters": {
        "max_uses": 1,
        "max_results": 3,
        "search_domain_filter": [
          "formula1.com"
        ],
        "search_recency_filter": "month",
        "country": "GB",
        "search_language_filter": [
          "en"
        ],
        "max_tokens": 4096,
        "max_tokens_per_page": 1024
      }
    }
  ]
}'

Searches cost $0.01 each, plus model tokens. usage.server_tool_use reports the billed count in web_search_requests, capped at max_uses. The model can exceed this allowance; extra searches cost about $0.005 each.

Web search is powered by Perplexity. It currently requires a model served through Vercel; other routes return 400.

Set these options inside the search tool's parameters. All are optional; numeric values must be integers.

The tool's parameters.max_tokens bounds retrieved search context. The request's top-level max_tokens limits the model's output per round.

Search the news

Find this week's Federal Reserve coverage:

curl https://api.minirouter.sh/v1/chat/completions \
  -H "Authorization: Bearer $MINIROUTER_KEY" \
  -H "Content-Type: application/json" \
  -d '{
  "model": "zai/glm-5.3-flash",
  "messages": [
    {
      "role": "user",
      "content": "Find this week's news about the Federal Reserve. List each headline with its publisher, date and URL."
    }
  ],
  "tools": [
    {
      "type": "minirouter:news_search",
      "parameters": {
        "max_uses": 1,
        "max_results": 5,
        "search_recency_filter": "week"
      }
    }
  ]
}'

News search is powered by Exa through Vercel and needs a Vercel-served model. Each search costs $0.01 plus Exa's extra $0.002 at your usual markup, plus model tokens. usage.server_tool_use reports the billed count in news_search_requests, capped at max_uses. The model can exceed this allowance; extra searches cost about $0.007 each. web_search and news_search share a limit of 10 searches per request.

Use either search_recency_filter or a date range, not both.

The model sends a query and receives up to max_results articles:

{
  "results": [{
    "title": "Fed holds rates steady",
    "url": "https://www.reuters.com/markets/...",
    "publishedDate": "2026-10-01T00:00:00.000Z",
    "author": "Reuters staff",
    "text": "The first 2,000 characters of the article..."
  }]
}

publishedDate and author are included when the source provides them; the publisher is the URL's site. No matches return an empty results list. A failed or timed-out search returns {"error": "timeout", "message": "..."} to the model, which can retry or answer without it. Search results stay with the model; your response contains its answer, not the raw articles.

Streaming, limits and cost

• Limit datetime and web-fetch calls: max_tool_calls accepts 1–30, defaults to 5, and requires at least one server tool. After reaching the limit, MiniRouter asks the model to finish without further tool calls. This does not cap web or news searches.
• Reserve for fetches: each round holds enough credit for the remaining fetch allowance, bounded by max_uses and max_tool_calls; only successful fetches are charged.
• Allow for multiple model rounds: the model reads the prompt, then reads the tool results. Each round is billed separately; max_tokens limits each round, not the whole loop.
• Read the total: non-streaming usage and x-minirouter-cost-usd include all model rounds. For streaming totals, send stream_options: {"include_usage":true} and read the final usage chunk.
• Keep the connection open: use stream: true and curl -N for longer loops. MiniRouter sends keep-alives while tools run, then sends the completed answer as SSE events. Intermediate tool rounds are not streamed token by token. Handle error events as well as the final [DONE].

For example, add these fields to the datetime request:

{
  "stream": true,
  "stream_options": { "include_usage": true }
}

Run your own functions

Use a function tool when the work belongs in your app, such as querying an order or calling a weather API. A function definition describes the input; it does not upload or run your code.

curl https://api.minirouter.sh/v1/chat/completions \
  -H "Authorization: Bearer $MINIROUTER_KEY" \
  -H "Content-Type: application/json" \
  -d '{
  "model": "alibaba/qwen3.5-flash",
  "reasoning": {
    "effort": "none"
  },
  "messages": [
    {
      "role": "user",
      "content": "What is the weather in London?"
    }
  ],
  "max_tokens": 256,
  "tools": [
    {
      "type": "function",
      "function": {
        "name": "get_weather",
        "description": "Get the current weather for a city.",
        "parameters": {
          "type": "object",
          "properties": {
            "city": {
              "type": "string"
            }
          },
          "required": [
            "city"
          ],
          "additionalProperties": false
        }
      }
    }
  ],
  "tool_choice": "auto"
}'

Then complete the conversation in your application:

1. Read choices[0].message.tool_calls. With tool_choice: "auto", the model may answer directly instead.
2. Parse each function's arguments JSON, validate it, check permissions, and run your function.
3. Append the complete assistant message, including its tool_calls, to your original messages array.
4. Append a role: "tool" message for each result. Match its tool_call_id to the actual call ID and encode content as a string.
5. Send the updated conversation to the same endpoint and model to get the answer. Repeat if the model requests another call.

An illustrative tool result message (replace the ID and content with your result):

{
  "role": "tool",
  "tool_call_id": "call_id_from_response",
  "content": "18 C, cloudy"
}

On models that support tool choice, tool_choice: "required" forces a call and "none" disables calls. parallel_tool_calls controls parallel function calls where supported. Your app controls its own function-call limit; max_tool_calls is for MiniRouter's built-in tools.

Plugins

Plugins are optional processing steps, enabled through plugins rather than tools. They do not add functions for the model to call. You can include both plugin IDs in one request; omit a plugin or set "enabled": false to disable it.

Fit a long prompt

Add this field to a Chat Completions request with an explicit model ID:

{
  "plugins": [{ "id": "context-compression" }]
}

If the prompt is too long for the model's context window after allowing for output, MiniRouter removes conversation turns from the middle, then shortens long messages if needed. It preserves the system-message prefix and the first and latest turns from removal, but message content may still be shortened.

This can discard details; it does not create a summary. Keep essential facts in a short system message or the latest prompt, and keep your own complete history. A prompt that already fits is unchanged. Use a named model, not minirouter/auto, so MiniRouter can look up its context window.

Repair JSON output

Add these fields to a non-streaming request and ask for JSON in your prompt:

{
  "response_format": { "type": "json_object" },
  "plugins": [{ "id": "response-healing" }],
  "stream": false
}

The plugin can remove code fences and surrounding text, and close unfinished brackets in choices[0].message.content. It also accepts a json_schema response format; see structured output for model support.

Valid JSON and output cut off by the token limit are returned unchanged. Repair is best-effort, does not guarantee schema compliance, and does not run with streaming. Check finish_reason, parse the JSON and validate it in your app.