TironianYour own small model, learned from your work.
Tironian watches the work you already do with Hermes, trains a small model on the tasks you keep repeating, tests it on your own past sessions, and only then lets it answer. When it isn't sure, your usual model answers and you never notice.
Free and MIT licensed. The first report needs no keys, no GPU and costs nothing.
hermes plugins install angelraph/tironian
What does this do?
Hermes downloads the plugin from GitHub and adds it to your plugins. Nothing runs until you enable it with hermes plugins enable tironian. It installs no packages: the part that runs inside Hermes uses only the Python standard library.
Built on open pieces you already trust
- Hermes Agenthost
- Nous Portalteacher
- Qwen3student
- Unslothtraining
- llama.cpp · Ollamaserving
- ModalGPU
From your sessions to a model that answers, with every number from real runs.
Captions are in the video. Read the transcript.
See what you repeat before you spend a cent.
Enable the plugin and Tironian reads your existing Hermes history, read only. It groups your requests into recurring tasks and shows what they cost you, right inside Hermes Desktop.
Three things, done carefully.
It learns from what went well
Every step is recorded with secrets masked, and graded by what actually happened. Only good turns from your own model are learned from.
Capture and grading → 02ProveIt has to pass a test first
A small model per task, tested on sessions it never saw. It goes live only above an 85% lower confidence bound.
Training and the gate → 03AnswerIt answers, with a safety net
Every answer is checked before Hermes sees it. Anything off, and your usual model replies instead.
Routing and fallback →A loop, not a one-off.
Each week of normal use adds graded examples. Families that were too small become trainable, and models that failed their test get another chance with more data.
Why did it answer like that?
Every request, model step and tool result, masked and stored on your machine. The raw material for everything that follows.
Capture in detaildecisions captured 7 tool runs 6 5 ok 1 timeout tool schemas 20 values masked 3
Mark what worked. Flag what broke.
Deterministic signals from the turn itself, plus your own /tironian good and bad. When it is unclear how a turn went, it is not learned from.
good █████████████████ 37 unknown ███ 6 bad ██ 4
Find the tasks you repeat.
Folder names, file types and filler words are set aside, common synonyms are merged, and what is left is the shape of the task.
Families in detail"largest files in my Downloads folder"
"biggest files in Videos"
→ show the <n> largest files in my <name> folderA small model per task.
Qwen3 1.7B, LoRA, two epochs, loss on the answer only. The prompt it learns from is byte for byte what it will be served.
Training in detailbase unsloth/Qwen3-1.7B method LoRA r16, all linear layers output Q4_K_M GGUF, about 1.1 GB cost about $0.15 on a Modal L4
Ship only what passes.
Same tool, matching arguments, valid format. The pass rate's lower confidence bound has to clear your threshold before anything changes.
The gate in detailn = 41 held out passed 39 lower bound 0.84 needs 0.85 → stays out of your way until it does
Local first, never worse.
Shadow mode first, compared against your model on real requests. Then it answers, with every reply checked and a silent fallback.
Routing in detailshadow agreed 44 of 48 92% live answered 211 fallbacks 6 saved $4.12 this month
“Your services to me are beyond counting, at home and abroad, in my studies and my writing.”Marcus Tullius Ciceroin a letter to Tiro, the secretary this project is named after
Built by an engineer, for people who want control.
Private by default
Nothing leaves your computer until you start a training run, and then only to your own Modal account.
Never worse than today
Your model is called at most once per request, and any problem in Tironian falls through to it.
Standard formats
Training data in Nous's Hermes function calling format. Models as plain GGUF files you can run anywhere.
Yours to keep
MIT licensed. Export any dataset with one command and take your models with you.
Everything else, one click away.
Docs
Install, your first model, every command and setting.
Read the docs →Cost
Free plugin, cents per training run, and a calculator for your own numbers.
See the cost →Privacy and safety
Exactly what it reads, keeps, masks and sends.
Read the details →How it works
Grading, task families, training and the math behind the gate.
Go deeper →FAQ
Twenty-five plain answers, from "can it make answers worse?" to hardware.
Read the FAQ →Changelog
What changed, including the bugs found along the way.
See what's new →
Stop paying a large model
for the same small job.
Install Tironian, read your report, and train your first model when a task is ready.