Tracking TinyFish in Onaro Meridian

AI agents don't just call models anymore. They browse the web, log in to portals, fill in forms and pull data from sites that were never built for them. TinyFish is one of the platforms doing that work, and for many teams it's becoming a real line on the AI bill.

That's the pattern we see again and again: every new kind of agent work arrives as a new kind of cost, billed its own way, outside the model invoices finance already struggles to read. If nobody can say which agent spent it, nobody owns it.

Onaro Meridian exists to close that gap. Here's how it handles TinyFish.

How TinyFish bills

TinyFish meters two things: Agent runs, priced per step, and Browser sessions, priced per minute. Search and Fetch are free. Usage is drawn from a prepaid Wallet, or billed under an enterprise contract.

Per-step pricing is simple to read on a pricing page and hard to manage at scale: a few hundred runs a day across a dozen agents adds up, and the bill doesn't tell you which agent did what.

The route that works today: through LiteLLM

LiteLLM has a built-in pass-through for the TinyFish Agent API. Point your TinyFish calls at the proxy, and it records every completed run with its cost (steps times the per-step rate, including a custom contract rate if you set one) against the LiteLLM virtual key that made the call.

Meridian already receives those LiteLLM records. When your TinyFish Agent runs go through LiteLLM:

• Each completed run is attributed to the agent you mapped to that virtual key. Set that agent's department and owner on Agents & allocation. Department reporting follows the agent.
• That spend sits next to your model spend in the same ledger: budgets and alerts, chargeback by department, and the journal entries Meridian posts to your general ledger.
• Nothing new to install on Meridian. If Meridian already reads your LiteLLM proxy, these runs come along. Setup takes a few minutes: see the TinyFish guide.

What it doesn't cover

• Browser sessions don't pass through this LiteLLM route, so per-minute browser usage isn't included.
• Meridian shows the dollar cost of each run, not its step count.
• LiteLLM records a cost for a completed run. A pass-through run with no cost is not stored.

Not on LiteLLM?

Two more routes:

• Report it from the agent. Agents can send their own usage to Meridian as OASA records, with a Usage ingest key. It works with any vendor and gives you per-agent attribution.
• A direct TinyFish connection. TinyFish's API exposes runs, browser usage and contract rates. We're working through key-level attribution and enterprise contract rates before we build it. If you're a TinyFish customer and want it, tell us: real demand sets the order.

The bigger point

Web agents won't be the last new line on the AI bill. Every time agents gain a new capability, a new vendor and a new meter come with it. The answer isn't a new dashboard for each one. It's one system of record that takes every source in, ties it to an agent and an owner, and books it the way finance already works.

That's what we're building.

See how it works for your setup: read the TinyFish guide or start an Agent Spend Assessment.

---

TinyFish is a trademark of its owner. LiteLLM is a trademark of BerriAI. Onaro is not affiliated with or endorsed by either.