Hermes Steps Past Its Threshold: Home Models, Zero Keys

A threshold is the doorway you step through to enter a room, and that image carries this whole release. A fresh desktop install no longer opens on a wall of settings. It opens on a chat that asks your name, your color, which apps you use, then suggests up to four first tasks. The first conversation runs on a trial model called nous/welcome, once per system user, with no key and no account. The door opens before anyone asks you for a card.

I remember the old wall well. You downloaded the app, opened it, and a key field stared back. That key is a long paid password that rents an intelligence engine such as Claude or GPT, and fetching one meant opening another site, registering, adding a card, copying a long string into the right box. Most people quit right there. The technically curious stayed; the normal user left .

Versions that run on your own computer

The part that surprised me most runs at home . One click now installs a local engine, llama.cpp, and the setup measures your memory and processing power, scores the catalog, then recommends the build your computer can actually carry. Downloads pause and resume; the engine starts with Hermes and stops with it. The list holds Qwen3.8 27B, Qwen3.6 35B-A3B and DeepSeek V4 Flash, plus any GGUF pattern from the HuggingFace archive. Nothing you type leaves the computer, while heavy work still goes to hosted models.

Connectors are the second doorway. Gmail, Linear and Notion now need a single browser sign-in each, and their tools work inside the same chat . The first connection quietly creates a free connectors-only identity, so no full account is needed. Two accounts of one app, work and personal, sit side by side, and you pick per task. The new Connectors page gathers hosted apps and community tool servers on one screen , replacing the old separate tab. For someone living across two inboxes like me, that screen breathes.

The feature most viewers will scroll past could save the most time : site sign-in. Small-business drudgery hides behind logins: ad accounts, store dashboards, analytics screens, supplier sites. The browsing agent can now sign in through your password vault, 1Password or the locked Hermes vault, and the model never sees the password; it fills without entering the chat. One-time codes travel the same hidden road. With no saved login, a hidden box appears for typing and saving. Before card numbers or addresses fill at checkout, the agent stops and asks first . A sample flow stays plain: open analytics, report last week most visited pages. What once needed a separate bridge is now part of the flow itself.

A routine that never drops long jobs

The extension catalog reaches 600 entries , each pinned to a reviewed version. You browse by kind and see the most adopted; this cycle drew 400 contributing makers . Everyday entries like Spotify and Home Assistant sit beside memory entries such as Honcho and Super Memory. The default install slimmed down, extras moved to the catalog, and you add one piece at a time. Starting small and widening feels right to me; the agent grows without bloating.

Long jobs gain three guards . First, continuity of thought: the agent reasons step by step, and when work pauses and resumes, or compresses to save room, that reasoning now survives by default on every model that can read it; it no longer forgets why it acted. Second, patience under load: when provider servers clog for an hour, the task no longer dies; the agent waits through visible cycles and continues . Third, the bot screen: each server agent gains its own desktop, you watch from the app, and when a step needs a human you take over, finish it, hand back. These stack onto the timed summaries with memory from the prior release; a 9 a.m. briefing knows what it told you yesterday and never repeats itself. Assigning a weekly job and going for a walk stops being fiction .

Sturdy releases and a large window

The release discipline matured too. Installs now track tested stable versions by default, with a separate channel for those chasing every fresh change. Moving to an older version counts as a rollback, labeled clearly, never silent. The desktop ships as a sealed signed package carrying its own runtimes, so first launch builds nothing ; Windows users gain a Microsoft Store path. Underneath sits one package manager, hermes pm, pinning tools and Python environments, switching only after the new side passes its check. Half-applied updates stop stranding users. A next-day 0.22.1 polished the Store path and placed a curated list atop the version picker.

I read the numbers through the GitHub release notes. Dated October 13, the window counts about 24,200 changes, 7,180 closed records, 2,520 contributors since the previous line. The Tao compilation frames it as 2.5 times the prior window, with 6,080 merged requests and 2,045K added lines. The AICoder summary gives the other face: the tree fell from 1.06M lines to 698K, llama.cpp b11370 installs in one click, updates dropped from 32.5 seconds to 16.5. The window grew while the tree slimmed by a third; cleanup happened inside the build. NousResearch ties the name to the old gatekeeper of doorways in its official line; crossing the threshold leaves setup troubles behind. The narration also names new model versions, but I anchor on figures, not names.

My read is plain : whoever starts now compounds with every version, whoever waits restarts from zero later. The basics stay constant: install, connect, hand over long work. What you learn keeps value. The StandardCompute billing guide confirms the shape: the trial version is a one-time courtesy, the 2,100-token trim and hourly retention stay opt-in habits. Routine work suits the home model; hard steps belong to hosted ones. I read this release less as a threshold than as a hinge; quiet doors admit crowds .