Jev Treg names the two halves of a decision-driven agent stack: treg is the open-source OpenRouter for agent tools that fetches anything — 2,896+ data and tool endpoints on one token — and Jev is TypeSafe's decision model that judges what treg fetched, in under a second per verdict. Fetch, then decide. That is the whole architecture. The rest of this guide is the detail: what each half does, what treg's own page says it costs, and how builders are already chaining the two together.
📺 Watch: Jev AI: 10 INSANE Use Cases
🔥 Get the Agent OS as a free bonus: AI Profit Boardroom members get the full Agent OS zip, prompt libraries, daily tutorials and weekly live coaching calls. → Get inside
Everything below comes from the treg.to/jev page, fetched 25 September 2026, plus treg's public repo and site. Treg's numbers are treg's own; community numbers belong to the builders credited by handle. New to the Jev half? Start with what Jev actually is, then come back for treg.
What treg Actually Is
Treg calls itself the OpenRouter for agent tools, and the analogy holds. One base URL, one token, and a catalog of 2,896+ endpoints across roughly 60 providers: X, LinkedIn, email verification, person and company enrichment, SEO, ads, scraping and more. Your agent does not hold a LinkedIn key, a scraper key and three enrichment keys. It holds one treg token, searches the catalog for the job it needs done, compares providers and prices, then calls the winner.
Pricing is per call, from about a cent, with no per-provider signups. That detail changes agent design more than it sounds: the agent can shop the catalog at runtime instead of you wiring every integration in advance.
Three more treg facts worth knowing before you touch it:
- Open source and self-hostable. Treg ships under AGPL, runs live at treg.to, and is built by the Superdesign team. You can audit the treg repo before you trust it with traffic.
- Team keys. Teams register their own keys, skills and CLIs on the server once, and every teammate's agent can call them without credentials ever leaving the server. That quietly solves key sprawl, the messiest problem in team agents.
- Serving detail. Per treg's own footer, Jev on treg is served through the Vercel AI Gateway — useful when you are tracing latency.
Jev Treg: The Division of Labour
The Jev Treg pairing works because neither half pretends to be the other. Treg fetches; Jev decides. Treg's page is unusually blunt about what its decision model can and cannot do, so here it is straight.
| Jev CAN | Jev CANNOT |
|---|---|
| Pick an action from your options | Write a sentence — it outputs decisions, not prose |
| Classify anything phrased as a question | Explain itself — the distribution is the explanation, so log it |
| Score and rank whole lists | Write code — though it can judge code: risky change, needs review |
| Answer yes or no with probabilities | Reason step-by-step — you write the steps as questions |
| Route on confidence | Read past roughly 30k tokens — chunk, or summarise then judge |
| Read your JSON as-is | Fetch anything — that is treg's job |
| Remember the last call — stateless, so put history in the state |
In treg's runs, that trade buys decisions at 5-7× a chat model's speed and 5-6× cheaper — roughly $20 per million decisions. The moment you stop asking a chat model to moonlight as your classifier, the economics of the whole pipeline change shape.
This decision-layer thinking is exactly what we work through inside AI Profit Lab — 3,000+ members building agent pipelines, $69/month locked in while it normally runs $110. If you want working sessions instead of blog posts, that is the room.
The Race Demo and the Cost Tables
Treg's page opens with a race: one support-ticket routing question sent to Jev and to a chat model at the same time. Jev returned technical at confidence 1.00 in 0.38 seconds, with zero output tokens, for $0.000018 — while the chat model was still writing its answer. A routing decision does not need prose; it needs a verdict and a number.
Treg's cost tables against GPT-5.6 Luna show Jev 1.7-5.8× cheaper and 4.5-7× faster across 2k to 32k token inputs. Credit where due: the same tables show Jev going over-limit past roughly 32k tokens. They publish the limitation next to the wins, which buys trust for the rest.
Still, these are vendor benchmarks on a vendor page. Treat them the way I treat every model claim on Goldie Bench: promising until replayed on your own data. Decision latency and cost per verdict are numbers you can verify in an afternoon.
The Flagship GTM Recipe, Walked Through
The most useful thing on treg's page is a full go-to-market recipe with every cost printed: find relevant LinkedIn commenters, let Jev qualify them, let treg enrich only the winners. Treg is upfront that its demos replay real runs with emails replaced — real numbers, anonymised output.
| Step | Volume | Cost |
|---|---|---|
| Search LinkedIn posts (treg) | 420 posts | about $0.004 each |
| Jev on-topic filter, kept at 0.70+ confidence | 203 kept of 420 | about $0.00003 each |
| Pull engagement on winners (treg) | 368 people | in the treg total |
| Jev ranks every person | 368 in 42.8s | $0.013 |
| Email lookup (treg) | 45 found of 56, an 80% hit rate | about $0.02 per hit, misses free |
| Total | 132.8 seconds | treg $0.870; Jev about a quarter-cent on filtering |
The output lands pre-bucketed: 56 decision-makers at a fit score of 1.5 or higher and 24 user-fit prospects at 0.8 or higher. Under a dollar, under three minutes, and the enrichment spend only went to people a decision model had already scored. Qualify before you enrich — that inversion is the entire recipe, and it is the same maths behind how to make money with Jev. Not in GTM? The page lists a buyer signal recipe, a signup lead scorer, a viral posts monitor and a build-your-own path.
Question-Craft: How You Steer Jev
Because Jev cannot explain itself and cannot reason in steps, the craft moves into how you phrase each decision. Treg's steering guide describes three levels, all plain configuration rather than code:
- Plain-string criteria. One line of text saying what to judge. Fast to write, good enough for coarse filters.
- Instructions with a focus line. You pass instructions as a small object whose focus line sharpens the judgment — treg's example focus: route to whoever must fix the root cause, not whoever the ticket happens to mention.
- Options as rubrics. Each option carries a what, a not-for, and examples. The not-for line is the trick — it is what stops near-miss misroutes, because the model sees where each option should lose, not just where it wins.
Most bad Jev results are really bad questions, and the fix is a sharper rubric, not a bigger model. I collect these patterns in Jev automation patterns if you want the full set.
Confidence Routing: The Safety Pattern
The confidence distribution Jev returns is not decoration — it is your routing layer. TypeSafe's cookbook, cited on treg with model jev-1.12, carries the canonical example: a melatonin question scored medical at 0.55, which fell below the auto-answer line and was routed to human review. Two thresholds in your code — act above one, escalate below the other — and nothing gets re-run. Uncertain cases surface themselves, which is exactly the behaviour you want before a decision model touches anything that matters. Wiring this into a full loop is covered in the Jev AI agent guide.
The Community Demo Wall
Treg showcases builders running the stack, and the numbers stay with the people who produced them:
- @borjafat mapped a 586-page linkmap in 45.1 seconds for $0.21 — the comparison run on Opus 5 got through 21 pages for $1.43.
- @venturetwins ran a Zillow search where thousands of listings were classified by filters nobody built — architecture style, renovation status, freeway proximity — in under 20 seconds for $0.18, with a pg-jev Postgres angle for doing it inside the database.
- @bryantchou built self-picking pages: 25ms per page load, Jev choosing copy and design per section for each audience segment.
- @RBilgil shipped a real-time slop detector.
- @chetaslua ran a live debate BS meter: 1,191 Jev calls at a 415ms median, $0.0497 total.
- @nutlope built fraud triage: Jev classified 100 emails in 1.42 seconds and routed the uncertain ones to Kimi K3 — 96 of 100 correct for about $0.07.
Every demo shares one shape: cheap bulk decisions first, with the expensive model or the human reserved for wherever confidence sags.
Where This Fits Your Stack
Full transparency: I have not run treg on my own pipelines yet, so you are getting treg's numbers, not mine. But the shape — a decisions layer riding on a data layer — is exactly the direction my Agent OS is heading, which is why this stack deserves your attention now rather than later. The Agent OS guide shows the architecture I do run; treg maps onto the fetch layer, Jev onto the judge. Want hands on the model first? Here is how to get access to Jev — and jevplayground.com is the zero-setup place to poke it before you wire anything.
If you would rather map this onto your own funnel directly, book a free strategy session and bring your current pipeline costs. We will find where a two-cent decision replaces a two-dollar one.
Jev Treg FAQ
Is treg free?
The code is: treg is open source under AGPL and self-hostable, so you can run it on your own box. The hosted service at treg.to prices per call, from about a cent, with no per-provider signups listed on its page — you pay for calls, not seats.
Does treg replace Clay-style tools?
Different category. Treg is a catalog gateway — one token, 2,896+ endpoints, and your agent composes the workflow across them. Enrichment platforms sell the workflow itself, with a human driving from a UI. If an agent is your orchestrator, a gateway fits; if a person runs the process by hand, a platform may still earn its seat. Judge by who does the orchestration in your stack.
What data can treg reach?
Per its own catalog: X, LinkedIn, email verification, person and company enrichment, SEO, ads, scraping and more — roughly 60 providers behind 2,896+ endpoints, priced per call so you only pay for what your agent actually uses.
Next Step
The Jev Treg stack collapses to one line: treg fetches anything, Jev judges everything, your code routes on confidence. The builders getting results are the ones running these numbers on their own pipelines this week, not next quarter. Do it with us inside AI Profit Lab — 3,000+ members shipping agent systems — or grab a free strategy session and we will pick the first decision in your business worth automating at twenty dollars per million.
