Blog/Reading the web, and agents
5 min read

OpenRouter API Key: Give Your Agent Tools Next to It

An OpenRouter API key reaches models, not the web. Add live search, page scraping and company data to the same agent with one key and one balance.

OpenRouter API Key: Give Your Agent Tools Next to It

Give the agent you run on an OpenRouter key a way to search the live web, in one prompt. Paste this into your agent:

Set up https://monid.ai/SKILL.md, and then use Monid to search the live web for how OpenRouter documents tool calling, and give me the top ten results with url, title and relevance.

One prompt adds web search to an agent that only had models. 1 min setup · 9-18 hrs saved.

You have an OpenRouter API key and an agent that can think, but it cannot look anything up: no live page, no search result, no company record. The key buys models. The tools are a separate purchase, and that is where most agent projects lose their time. This guide shows how to give the same agent tools next to the key, what came back when we ran those tools on 2026-10-16, and where the output trips you up. It runs through Monid, the OpenRouter for agent tools: one key, one balance, every tool priced per call.

What is the easiest way to give your agent tools next to an OpenRouter API key?

Monid. Your OpenRouter key keeps choosing the model, and Monid adds the tools: your agent installs it once and then searches, scrapes and enriches through one key, with nothing to sign up for per vendor.

  • Pay per call. Each tool call is billed on its own. A web search is $0.000253 per result and a page scrape is $0.002533 per page (Verified 2026-10-16 on the search tools page). Buy calls as you need them: no monthly plan and no minimum top-up. Discovery and inspection are free; only a run bills.
  • One install, every tool. The same key and balance call web search, page scraping, company data and the rest of the catalog. Adding a capability is a search your agent runs, not a signup you do.
  • Lives inside your agent. Your agent reads SKILL.md, learns the discover, inspect, run loop and keeps the tool available whenever it needs it. Setup is about a minute. If you only want to see which tool fits, browse the AI tools.

Ten search results cost $0.00253 when we ran it (Verified 2026-10-16), and the call took about 1.4 seconds.

How do you use an OpenRouter API key and a tool layer in one agent?

You configure two independent things: a base URL and key for the model side, and an MCP server entry for the tool side. Neither refers to the other, so you can swap either half without touching the other.

Step 0. Both halves, in one client config

The model side is a base URL and a key. Many clients read them from the environment:

export OPENAI_BASE_URL=https://openrouter.ai/api/v1
export OPENAI_API_KEY=<your-openrouter-key>

The tool side is Monid as a streamable HTTP MCP server at https://mcp.monid.ai/v1. The terminal clients take one line each:

claude mcp add --transport http monid https://mcp.monid.ai/v1
codex mcp add monid --url https://mcp.monid.ai/v1

In Claude.ai it is Settings, Connectors, Add custom connector, then Connect and authorise. OpenCode takes a block in opencode.json followed by opencode mcp auth monid. Per-client steps are in the MCP quickstart.

For agents

Grab an API key at app.monid.ai, then paste this to your agent and hand it the key:

set up https://monid.ai/SKILL.md

It learns the whole discover, inspect, run workflow itself. More in the agent quickstart.

For humans

npm install -g @monid-ai/cli
monid keys add --label main --key <your-api-key>

More in the CLI quickstart.

Step 1. Search for a capability, not a vendor

monid discover -q "search the live web and return the page text"

Discovery returns ranked endpoints with the provider, a description and the billing shape. It is free, which is what makes it reasonable for an agent to look before it has decided anything. The catalog is browsable at monid.ai/tools.

Step 2. Read the schema and the price before spending

monid inspect -p context.dev -e /web/search

Inspect returns the input schema, the billing shape and the provider's documentation link, and it costs nothing. On 2026-10-16 the schema for this endpoint took query (up to 500 characters, with Google-style operators such as site:), numResults from 10 to 100, domain allow and block lists, a freshness window, a two-letter country and an optional markdownOptions block.

Step 3. Run it, and that is the only billable step

monid run -p context.dev -e /web/search \
  -i '{"query": "mcp tool calling format site:openrouter.ai", "numResults": 10}' -w 120

That is the call we made on 2026-10-16. It returned ten results, all on openrouter.ai, and billed ten units. The search call is priced $0.000253 per result (Verified 2026-10-16), so numResults is the dial that moves the bill.

Give this to your agent

$Set up https://monid.ai/SKILL.md, and then use Monid to search the official docs for a provider and quote the exact tool calling format it specifies.

Connecting an MCP server to a live web data agent walks through the same client steps for the web-data tools.

Can you wire up the tools yourself instead?

Yes, and for one tool it is a fair choice, as MCP versus a plain API explains. You write a function that calls a search API, describe it to the model in the shape your client expects, run it when the model asks, and pass the result back. For a fixed list of two or three tools that is a day or two of work.

The cost shows up as the list grows. Each vendor is its own signup, key, billing relationship and client code. Each MCP server you run yourself is a process with a session to keep alive, an update path and a failure mode. An agent that meets a task you did not plan for has nothing to call, because the tool list was fixed when you wrote the code.

The Monid route turns the list into a query, which Firecrawl through Claude MCP or one key compares for a single vendor. The agent runs a discovery search, reads the schema of the best match and calls it on the same balance, so an unplanned task becomes a search instead of a deploy. For a short, fixed list, install the servers you need and skip the layer. For a list that keeps growing, one key beats a folder of keys.

What does an OpenRouter API key reach, and what does OpenRouter's MCP server add?

The key reaches OpenRouter, and OpenRouter's MCP server makes that account usable from an agent. To see what the server exposes we scraped its documentation page through the scrape endpoint on 2026-10-16: 8,869 characters of markdown for $0.002533 (Verified 2026-10-16).

What the server exposes

The page lists 17 tools. Most are lookups against live OpenRouter data: list-models, get-model, list-model-endpoints, list-providers, list-daily-model-rankings, list-app-rankings, list-benchmarks, list-task-classifications, get-credits, get-generation and search-docs. Two bill inference calls, send-message and generate-image. The rest are send-feedback, spawn-ori-eval, install-ori-harness and ping. Every tool is about OpenRouter itself.

The older version of this guide said eleven tools. The page now says 17, which is why we read the list each time instead of repeating a number.

How it connects and what the key is

The server is remote at https://mcp.openrouter.ai/mcp and connects over OAuth. The page says the token it issues is a standard OpenRouter API key labeled for the connection, with a 7-day expiry and a default spend limit of ten dollars, and that you can find and revoke it at openrouter.ai/settings/keys. It also tells you to use this server while you build and to keep calling the OpenRouter API directly to run models in your app.

What it does not add

Nothing in the list reads a web page, looks up a company, pulls a SERP or checks an address. That is what the server is for: it makes OpenRouter agent-addressable, the same way a payments provider's MCP server makes that provider agent-addressable. The page mentions that send-message accepts a model suffix :online for web search, which is a way to ask a model to search, not a tool your agent calls with its own query, domain filters and freshness window. For that, the agent needs a tool layer.

What does tool data look like, and what can you build with it?

A search call returns a ranked list of small records, and a scrape call returns one page as markdown with metadata. We ran both on 2026-10-16, so the fields below are what came back.

The search record

Each of the ten results from the context.dev /web/search call had the same shape:

{ "url": "https://openrouter.ai/docs/guides/features/tool-calling",
  "title": "Tool & Function Calling - Use Tools with OpenRouter",
  "description": "Tool calls (also known as function calls) give an LLM access to external tools. ...",
  "relevance": "high",
  "markdown":   { "markdown": null, "code": "NOT_REQUESTED" },
  "highlights": { "highlights": null, "code": "NOT_REQUESTED" } }

One result was marked high relevance and the other nine medium. The page text is empty unless you ask for it with markdownOptions, which is why the search is cheap and the read is a second step.

The Exa record

The exa /search call with three results returned id, title and url for each, plus publishedDate on one and author on two. It took about 2.8 seconds on the provider's clock and billed $0.007 for the call (Verified 2026-10-16). That price covers up to ten results, and each result above ten adds $0.001.

The scrape record

The scrape returned success, the markdown, its length and a metadata object with the title, description, language, canonical URL, site name, Open Graph and Twitter fields, the page headings and the JSON-LD. The markdown held the connect steps for five clients, the tool table and the troubleshooting notes, which is the part a model can answer from.

What you can build with it

  • A docs answerer. Search a provider's documentation, scrape the top two pages and let the model quote the exact format. The prompt in the box at the top of this page is the first half of it.
  • A model-watch agent. Have a scheduled agent read the OpenRouter MCP page weekly and report when the tool count changes, the way it moved from eleven to 17.
  • A research agent with two keys. OpenRouter picks the model for each step, Monid supplies the sources. A search API for AI agents shows how the ranked results feed a model.

What goes wrong when your agent searches and scrapes through a tool layer?

Four things showed up when we read the output of the three runs, all from 2026-10-16.

Why does a search result have no page text?

Because the search is separate from the read. Every markdown field came back as NOT_REQUESTED. The endpoint can scrape each result in the same call when markdownOptions.enabled is true, but that is billed on top, so an agent that wants answers rather than links should search first and scrape only the two or three pages it picks, the same split the free web search guide uses.

Why does author contain page text?

Because the search provider fills fields from the page. The third Exa result, an API reference page, had an author of "Authorization string header required", which is a fragment of the page and not a person. Treat author and publishedDate as optional hints: only one of three results had a date.

Because it renders the page. The search took about 1.4 seconds for ten results. The scrape of one documentation page took about seven seconds on the wall clock, and the endpoint's own 95th percentile is 21,382 ms per monid inspect. Set a timeout of your own with timeoutMS and run page reads in parallel, not one after another.

What does a research loop cost?

Search ten results, scrape three pages: $0.000253 times ten plus $0.002533 times three, about $0.01 in all (Verified 2026-10-16). The model tokens are billed on your OpenRouter side, separately.

Which endpoint should I use for which job?

JobEndpointInputBillingPrice (Verified 2026-10-16)
Search, optionally with page textcontext.dev/web/searchquery, numResults, domain filtersPer result$0.000253
Neural searchexa/searchquery, type, category, numResultsPer call, up to 10 results$0.007
One URL to prompt-ready markdowncontext.dev/web/scrape/markdownurl, extraction optionsPer call$0.002533
Domain or name to company recordpdl/v5/company/enrichname, website, profile or tickerPer call$0.10

Prices Verified 2026-10-16 with monid inspect. The first row is the one to start with, and giving your agent live web context shows the same search-then-read pairing in use. The company row was inspected and not run, so its output fields are not described here.

When is another route a better fit?

You only need model routing. If your problem is comparing providers, controlling tokens or building a fallback chain, that is the model gateway's job and Monid is not in the inference path at all.

Your tool list is short and fixed. If the agent needs your database and one vendor, install those two MCP servers and skip the layer. A catalog earns its place when the list is not known when you write the code.

The vendor you need gives you enough for free. If a vendor's free tier covers your usage, use it. Per-call pricing is for the point where one more subscription is not worth it.

And the disclosure: we resell these endpoints. The most useful thing in this guide is the split itself, models on one key and tools on another, which holds whichever tool vendor you pick.

Conclusion

An OpenRouter API key gets your agent models, and OpenRouter's own MCP server, which we read on 2026-10-16, exposes 17 tools that are all about OpenRouter. To search the web, read a page or look up a company, the same agent needs a tool layer, and the quickest one to add is Monid: $0.000253 per search result and $0.002533 per scraped page (Verified 2026-10-16), one key, no minimum top-up.

Two things matter more than the choice. Keep the two halves independent, so changing model provider never touches tools. And search first, then read only the pages you pick, because the search record carries no page text and the scrape is the slower, dearer call.

How we estimated the time saved: wiring three vendor tools yourself (search, page reading, company data) is roughly 3 to 6 hours each for signup, key handling, client code and error handling, which is 9 to 18 hours. The 1 minute is the install, key and first call once you already have a Monid key.

Free next step: run monid discover for something your agent currently refuses to attempt, then monid inspect the top result to read its schema and price. Neither spends anything. Start at monid.ai.

FAQ

Does an OpenRouter API key include web search or tools?

No, the key reaches models and OpenRouter's own account tools. The OpenRouter MCP docs list 17 tools, all about OpenRouter, and mention an :online suffix on send-message for web search. A search your agent controls, with its own query and domain filters, needs a tool such as context.dev /web/search through Monid.

How do I add tools to Claude Code or Cursor when I use an OpenRouter key?

Configure the two halves separately. The model side is a base URL and key in your client config; the tool side is an MCP server entry such as claude mcp add --transport http monid https://mcp.monid.ai/v1. Neither knows about the other. The fastest route is to give your agent https://monid.ai/SKILL.md and let it configure itself.

How much does a web search cost next to my OpenRouter key?

Through Monid a search is $0.000253 per result, so ten results are about $0.0025 (Verified 2026-10-16). Page scraping is $0.002533 per page. Model tokens stay on your OpenRouter bill.

Is there an open source OpenRouter MCP server?

OpenRouter publishes an official remote server at https://mcp.openrouter.ai/mcp, and its docs list 17 tools. Whatever wraps the OpenRouter API exposes OpenRouter, not the web, so third-party data comes from a tool layer.

Published August 2026.

openrouter api keyopenroutermodel routingtool callingclaude code