Jev for Hermes Agent owners: 8 jobs it does

On 3 October 2026 the official plugin catalog of Hermes Agent listed 13 community add-ons with Jev in their name, all added since 18 September, according to the catalog folder on GitHub. Jev is an AI model from TypeSafe AI that reads a piece of text and answers one small question about it. OpenClaw gives it a slot of its own. Hermes Agent takes it only through add-ons, and Nous Research, which makes Hermes Agent, has published a test of its own.

What Jev is, in everyday words

Jev writes no text and holds no conversation, according to TypeSafe's page on coding agents. It answers one small question about the text it is given, in one of three shapes listed on TypeSafe's introduction page: yes or no with a number for how sure it is, one pick from a list the owner defines, or a score on a scale the owner describes. TypeSafe AI of San Francisco opened early access with a waitlist in a post dated 15 September 2026, and its homepage still says early access.

How it gets into Hermes Agent

OpenClaw ships a slot for a decision model, as the sister article for OpenClaw owners describes, while the latest Hermes Agent release notes, dated 24 September 2026, carry no mention of Jev. Jev arrives through add-ons written by people outside Nous Research and listed in the official plugin catalog, where Nous reviews each entry and pins it to one exact version, according to the catalog rules. The steps, from the catalog page of one such add-on:

1. Get an API key, a long password that lets your server talk to TypeSafe, from the TypeSafe console. Some add-ons also take a key from OpenRouter, a service that resells many AI models under one account.
2. Install the add-on from the catalog, where each page has a button that opens it in the Hermes desktop app.
3. Give the installer the key. The agent then has extra decision tools it can call, and nothing in Hermes Agent is replaced on its own.

Every call sends the question, the text being judged and your key to TypeSafe, the same page states. A public request of 30 September 2026 asks Nous for such a slot and is still open.

What it costs

TypeSafe's price list shows $0.042 (€0.037) per million input tokens, and answers are free. A token is a small piece of text, roughly a short word. The euro figures here use the European Central Bank's reference rate of 2 October 2026, 1 euro to 1.1225 dollars.

Eight jobs for Jev in a Hermes Agent setup

All view and star counts below are from 3 October 2026.

1. Approving risky commands

When Hermes Agent is about to run a command that needs a check, Jev decides approve, deny or ask the person, a call that otherwise goes to the chat model, the AI that writes the agent's replies. hermes-jev-approvals by anpicasso, published on 17 September 2026 and listed in the Hermes catalog, does this. By the author's own measurement on 153 real commands it answered 8.7 times faster and sent 4.4 times fewer questions to the person.

2. Picking the one skill a request needs

Jev picks the one Hermes skill, a saved recipe for a kind of task, that a request needs. TypeSafe's own recipe page tested this on the 182 skills of the Hermes skill catalog with 488 test requests: by TypeSafe's own measurement wrong skill loads fell from 16.8% to 7.3% and needless loads from 9.8% to 4.0%. typesafe-skill-router by DECRUX9812, published on 16 September 2026, does it for about $0.001 (€0.0009) per routed turn, one message and its answer, by the author's count.

3. Choosing the cheapest model good enough for each message

Jev labels each message light, standard or heavy, and Hermes Agent answers with a cheaper or a stronger model, one of OpenAI's GPT-6 Luna, Sol or Astra. hermes-jev-router by HinkoK was published on 2 October 2026 with a video the same day. By the author's own measurement a decision costs about $0.00006 (€0.00005), about $1.80 (€1.60) for 30,000 messages, and the labels were 95.8% to 97.5% correct on a control set of 120 messages.

4. Skipping scheduled runs that would do nothing

Before a scheduled Hermes job wakes the main model, Jev answers whether running now would do useful work, and a confident no skips the run. hermes-jev-cron-gate by superlowburn, published on 25 September 2026 and listed in the Hermes catalog, does this. By the author's own measurement a check takes about half a second, and the agent is woken anyway after four skips in a row.

5. Sorting an inbox

Raveesh Bhalla posted on X on 19 September 2026 that his Hermes agent sends transaction emails to Jev, which sorts them and flags the tax deductible ones. On 1 October 2026 Thalena Lima published a video in Portuguese, 5,801 views, in which Jev sorts a demo Gmail inbox and Hermes writes the report and sends it to WhatsApp, running around the clock on a rented server.

6. Sorting comments on YouTube

Hermes Agent pulls the comments, Jev puts each one in a group, needs a reply, unsure or no reply needed, and Hermes sends a report to Telegram twice a day. Emrullah Yaprak showed the setup in a video in Turkish published on 27 September 2026, 4,575 views.

7. Screening web pages for hidden orders

Some web pages carry hidden text that gives an agent orders, a trick called prompt injection, and Jev flags those parts before Hermes Agent reads a search result or a page. hermes-jev-skills by kerpopule, published on 18 September 2026, had 1,000 stars on GitHub, the site's bookmark count, the most of any Hermes and Jev project. By the author's own measurement, with attacks planted in 80 real web results, 70 of 79 were caught, none of 1,520 clean pieces were held back and a result took about 0.2 seconds. The author notes the attacks were of the kind the add-on was built for.

8. Deciding when Hermes Agent saves a lesson

Hermes Agent runs a review that writes memories and skills, normally on a turn counter. With hermes-jev-lesson-gate by VBS2004, published on 30 September 2026, Jev asks after each turn whether the person corrected the agent, shared a lasting fact or the agent found a fix, and the review runs only then. A Jev call costs about $0.00004 (€0.00004), up to two per turn, by the author's own count. The author writes that the questions were written after seeing the test labels and that a test on fresh examples is still to come.

What Nous Research itself did with Jev

Teknium of Nous Research, who merged every Jev entry in the official catalog, posted on X on 19 September 2026 that Jev for compaction, the shortening of a long conversation history, "has a lot of problems". The team's test file, in the main code since the same day, ends with the verdict "Do not adopt Jev" for that job: Jev was faster and cheaper, 1.4 seconds and $0.007 (€0.006) per shortening against 37 seconds and $0.061 (€0.054) for what Hermes Agent ships, but it kept about twice as much text and recalled slightly less.

On 21 September 2026 Teknium published a personal demo project in which Hermes plans and Jev picks one action at a time in Minecraft. By his own report it reproduced a known run that defeats the game's final boss, the Ender Dragon, in 7 minutes 45 seconds for $0.96 (€0.86) in total, of which Jev was about $0.01 (€0.01).

Where this leaves the owner of an agent

Numbers 4, 5 and 6 run around the clock, so they only do their job on a machine that stays on, which is what OneClickClaw runs for each customer: a Hermes Agent server of their own. The Starter plan is EUR 14.99 a month and the 7 day free trial needs no credit card.