We just shipped something new: _done (underscoredone.com), a growing set of tools built specifically for AI agents — no accounts, no API keys, no subscriptions. An agent finds a tool, pays per call, gets its answer, and moves on.
Here's why we built it, what it actually is, and how to try it today.
The problem we kept running into
AI agents are doing more and more real work — looking things up, verifying data, running calculations, fetching pages, converting formats. But the tools available to them are stuck in a human-first world.
Free tools in agent frameworks are often slow, inconsistent, or simply down. Paid tools exist, but they're gated behind a wall of human-shaped friction: create an account, verify an email, pick a plan, enter a credit card, wait for approval — then start using it. That workflow makes sense for a person. It makes no sense for an agent that just needs one answer, right now, and won't be back to manage a dashboard or pay a monthly bill for a tool it might call twice.
We kept seeing the same gap: agents need correct, on-demand answers, not another login form.
What _done is
_done is a catalog of small, fast, single-purpose tools, each built to do one thing well and built for machines to consume directly. The model is simple:
- Find the tool — through our directory, MCP, marketplaces, or any agent discovery surface.
- Read the metadata — every tool ships with an OpenAPI spec, an llms.txt, and an MCP-ready definition.
- Pay per call — in USDC, via the x402 protocol. No invoice, no subscription, no card on file.
- Get the result and move on.
There's no account to create, no API key to manage, and nothing sitting on the books whether you use it or not. As our founder Ori Tzvielli put it when we launched: "You ask, and it's ready — for you and your AI."
A couple of examples
A couple of the tools live today give a good sense of what _done is for:
- Hashing, HMAC & Checksum Suite — computes SHA-256, SHA-512, MD5, and other hashes, signs data with HMAC, and verifies signatures with a timing-safe comparison. This one exists because language models can't actually do cryptographic hashing — they generate a string that looks like a valid hash but isn't, and every webhook check or integrity check built on that hallucination is silently broken. An agent calls the tool and gets the real, correct digest instead.
- Mortgage & Loan Amortization Calculator — give it a loan amount, rate, and term, and it returns the exact monthly payment plus a full month-by-month schedule, including "what if I pay an extra $300/month" or "what if I make a lump payment in month 24" scenarios — computed with pinned rounding rules so the numbers are exact to the cent, not estimated.
Each tool is intentionally narrow. One job, done reliably, instead of a do-everything API that's slow and inconsistent at all of them.
Try it out
_done is live now at https://underscoredone.com with nine tools available and more shipping regularly — we're building at a clip of 5-10 tools a day. If you're building agents, working with MCP, or experimenting with x402 payments, you can point an agent at the directory and start calling tools today — no signup required, by design.
We're also looking for design partners, registry partners, and feedback from anyone deep in AI, MCP, or x402. If that's you, we'd love to hear from you.
Ask. _done.