Treg Is an 'OpenRouter for Agent Tools,' and It's Trending

There's no version tag on this project yet. The releases tab is empty, there's no changelog, and the team hasn't shipped a 1.0. And yet it sat on GitHub's daily trending page this week, picking up 468 stars in a single day on its way to a current total of 3,152. What pushed it there: 657 pull requests, against a repo with only 12 issues filed against it in its entire history.

The project is Treg, and its own README describes it in one line: "OpenRouter, but for agent tools instead of models."

What Treg actually does

OpenRouter made its name by putting one API key and one base URL in front of dozens of different language models, so a developer doesn't have to sign up with OpenAI, Anthropic, Google, and a dozen others separately. Treg applies the same idea to tools rather than models: one token, one base URL, and — per the project's README — access to "3,000+ catalogued endpoints across 60+ providers", spanning keyword and rank tracking, backlink and authority data, social publishing, people and company enrichment, ad management, scraping, and image/video generation.

The pitch, stated directly in the README, is aimed at the awkward economics of giving an AI agent real-world capabilities: tools like Semrush ($139/month), Moz ($99/month), and Crunchbase ($99/month) sit behind their own logins, their own billing, and their own signup flows — expensive and slow to wire up if you just want an agent to check backlink data once. Treg's answer is to let a team pay "per call, from a cent," using a prepaid balance instead of a stack of monthly subscriptions, with new verified accounts getting a one-time $1.00 in free credit when they set up an eligible team.

In practice, using it looks like this, straight from the README's quickstart:

curl -fsSL https://treg.to/install.sh | sh
treg login
treg catalog search "backlinks for a domain"
treg call tikhub.tiktok.user.profile --query uniqueId=tiktok

The mechanism: a relay, not a model

Treg's own description of its architecture is specific about what it isn't doing: it's a "faithful streaming proxy" that "relays, never models" the upstream services it fronts, injecting provider credentials server-side so callers never see them. That's a meaningfully different design choice than building a unified abstraction layer that normalizes every provider's response shape — Treg is explicitly not trying to paper over the differences between, say, a TikTok profile lookup and a Semrush backlink query. It passes the request through and gets out of the way.

That's also why it's Apache 2.0-licensed with one carve-out: you can self-host it, modify it, and use it commercially inside your own product with pass-through billing, but you can't stand up a competing hosted version of the registry itself without written permission from the license holder. The official hosted instance runs at treg.to, which the README says runs on Render.

Why it's the thing blowing up on GitHub this week

The traction numbers only make sense once you see how young this repo is. There's no tagged release, which usually means a project is still pre-1.0 by the maintainers' own reckoning. The commit history shows a wall of activity landing on a single day — September 24, 2026 — with commit messages like "fix(catalog): align Fetchin profile miss settlement" and "docs(web): teach X-Treg-Token on the /docs API page (#670)". Those read like the granular, provider-by-provider fixes you'd expect from a catalog that's still being built out in public.

The pull request queue tells the same story at a bigger scale: 657 pull requests total, 78 still open, against a project with only 12 issues ever filed. That's an unusual ratio — most projects accumulate issues faster than they close PRs early on — and it suggests a small, fast-moving team adding providers to the catalog largely through their own PRs rather than responding to a backlog of external bug reports.

The team isn't new to viral GitHub launches

Treg comes out of Superdesign, an organization whose flagship project is superdesign, described in its own README as "the first open-source design agent that lives right inside your IDE" for Cursor, Windsurf, Claude Code, and VS Code. That repo has picked up 7,008 stars on its own. The same org also maintains a Claude Code/Cursor design skill and a couple of smaller self-iterating-agent experiments — Loopany among them. Treg is a departure from design tooling into developer infrastructure, but it's coming from a team that has already shown it can turn a GitHub repo into a fast-growing star count once.

What's still missing

None of this makes Treg a finished product. There's no independent security review of a system whose entire job is holding other services' API credentials and relaying authenticated calls on your behalf. That's a reasonable thing to want before routing anything sensitive through it.

The LICENSE also means you can't build a rival hosted catalog on top of the open-source code without asking first. That caps how far the "self-host it yourself" option actually goes if what you wanted was to compete with treg.to directly. And with zero tagged releases, anyone adopting it today — hosted or self-hosted — is running against a main branch that changed dozens of times in the last 24 hours alone, not a version anyone has committed to supporting.

The takeaway

Treg is a good example of a pattern this category keeps turning up: a real, working piece of infrastructure — not a wrapper, not a demo — going from obscure to GitHub's front page within days. It's the same shape of story as Cloudflare, BrowserOS, and Soloop all launching agent-first browsers on the same Product Hunt day back in August, or i-have-adhd's three-week climb to #1 on trending earlier this month. The common thread across all of them, and across the memory layers agents are also racing to standardize, is that the bottleneck for agent-based products right now isn't model capability — it's plumbing.

Treg is a bet that the plumbing problem for tool access specifically is billing and credentials, not integration code. If you're already paying for three or four API subscriptions just so an agent can occasionally check one of them, it's worth reading the pricing section of the README before you renew the next one. But treat the lack of a version number as exactly what it is: an invitation to test, not to depend on it yet.