I Tried an OpenRouter for Agent Tools. It Cost 2.5 Cents.

My coding agent can research a company. It can also ask for data I do not pay for. Each new data provider used to mean a signup, a key, and a monthly plan. I tried a service that puts many providers behind one key. My agent made four calls. Two failed. Two worked. The bill was two and a half cents.

Here is what the service is, what happened, and where I would trust it.

A router puts many data tools behind one key

OpenRouter is a service for AI models. You open one account and get one key. That key reaches models from many companies, and you get one bill.

Treg does the same thing for the data tools an agent calls. Its homepage says so in four words: “OpenRouter for agent tools.” Think search rankings, company records, and contact lookups.

On September 21, 2026, that homepage listed 2,630 endpoints from 47 providers. An endpoint is one kind of question you can ask a provider, such as “which search terms does this site rank for?”

The pricing is the other half of the idea:

• You pay for each result from a prepaid balance. There is no monthly seat.
• A new account starts with $1.00 of free credit.
• If you already pay a provider, you can add your own key and route those calls through it.

A person rarely needs 47 data vendors. An agent might use a different one for each task. That is the case for a router: the agent asks for the job, and the router knows who can do it and what it costs.

Set the spending limit before the first call

An agent with a payment method needs rules. I set mine before anything ran.

1. The whole trial had a $5 ceiling.
2. Each call carried a header that capped its cost at $0.10. Four calls could cost $0.40 at most.
3. Each call carried an idempotency key. That is a label for one request. If the connection drops and the agent tries again, the server sees the same label and does not charge twice.
4. The scripts saved every response and skipped any call that already had one.

I signed in through the browser. The agent kept the key in a file that only my user account can read. It used the plain web API. It did not set up Treg’s command-line tool or add Treg to my agents’ settings. I never added a card.

Half my calls failed, and the failures were free

I picked one fintech company I was researching. I asked for two things: a company profile, and ten US search terms the company’s site ranks for.

Four calls are four observations. They are not an uptime number, and I would not judge any of these providers on one request.

The SpyFu failure was the most useful response of the day. It did not just say no. It said when capacity would reset. It named three other providers that answer the same question, each with a price: about $0.0005, $0.012, and $0.0179 per call. It also said nothing had been charged.

That is the router doing its job. An agent can read that response and pick the next step without me. Mine took the cheapest option, Serpstat, and it worked.

In The Failover Economy, I wrote about businesses that pay for a second AI vendor. This is the same move one layer down: a backup for each data source, chosen at the moment the first one fails.

The ledger matched to the fraction of a cent

Each paid call showed up in the account ledger twice. First Treg reserved the expected cost. Then it settled the real cost.

The two charges were $0.0245 and $0.0005. That adds up to $0.025. The balance went from $1.00 to $0.975, with nothing left on hold. The response headers, the ledger, and the balance all agreed. All of it came out of the free signup credit.

One note for anyone planning a budget. The Serpstat charge came in below what I expected from the catalog for ten rows. That is the good direction, but it means an estimate is not a receipt. Check the settled charge.

Cheap data still needs checking

The company record agreed with the company’s own website on what the company sells. It also added research leads: an employee count, 35 detected technologies, and 35 job titles.

Leads are not facts yet. The fields carried no dates, so I could not tell a current job opening from an old one. One of the company descriptions was labeled as written by AI.

Two fields were wrong, or at least misleading:

• The demo flag said no. The record said the company offers no demo. The company’s website shows several links to book one.
• The funding label said Series A. The company’s own launch announcement says the founders put in the money themselves and chose not to take outside funding. The amount matched. The label did not.

Either error would look bad in an email to that company. Both were easy to catch, because the company’s own pages were the check. The router made the record cheap. It did not make it true.

A second search source finds different things

I already use Semrush. So I asked it the same question: ten US search terms for the same site, sorted by search volume.

• The two lists shared one exact term: the company’s name. The site ranked first for it in both.
• Semrush reported a search volume of 2,400 for that term. Serpstat reported 2,900. Both numbers are estimates, not visits to the site.
• Both lists led me to the same how-to guide on the company’s blog.
• Serpstat also showed a broad industry term, with the site in third place and a reported volume of 12,100. It pointed to a real article on the site.

That last row is a lead I did not have. It is not proof that Semrush lacks the term, or that Serpstat’s rank and volume are right. I asked Semrush about two of Serpstat’s terms by exact match, and it found nothing for either. That tells me the two tools differ. It does not tell me which one is right.

One more difference matters when you cite a number. Semrush dated each row. Serpstat’s rows had no dates.

The calls took nine seconds. The work took longer.

The two calls that worked took about nine seconds together. Signing in, reading the catalog, choosing a fallback, and checking the results took far longer.

A router removes the signups, the keys, and the monthly plans. It does not remove the judgment. Someone still has to pick the right provider for the question and check the answer against a source that knows.

Where I would use it

• Yes: cheap extra leads, and a second opinion on what another tool reports.
• Still the authority: a company’s own pages for facts about that company.
• Still the baseline: the search tool I already know, with its dated rows.
• Not yet: any task that has to work on the first try. Two of my four calls failed.

If you want to try a router like this, treat your agent like a new hire with a company card. Set a small limit. Ask for receipts. Then read them.

The question is: what would your agent look up if each answer cost a fraction of a cent, and who checks it?