Hexi system card | Elovaren

What Hexi is

Software on your machine: Hexi is desktop software. It shows you the answer in cards, charts and live widgets, and it runs on your own computer. By default the model it talks to runs elsewhere (see Models, below).
Built by: Hex Infrastructure LLC, a subsidiary of Elovaren LLC.
Release: Pre-release, version 0.1. Some of what this card describes will change before general release.

Models

Default: Claude Opus 5.5, by Anthropic, with Claude Opus 5 behind it when Opus 5.5 cannot serve.
Where model requests go: Through OpenRouter on your own key, and only to U.S. companies’ endpoints with zero data retention (for Claude, Amazon Bedrock’s us-east-1 or Google Vertex’s U.S. endpoint), run by providers that neither keep nor train on what they are sent. OpenRouter’s own relay is not held to a region. A model with no such endpoint is refused rather than routed around the rule. HEXI_PRIVATE_ROUTING=off lifts this for a launch unless a policy holds it on. With a model in your own AWS account, model requests go to Amazon Bedrock there instead. Web lookups and the update check still go elsewhere (see What leaves the computer, below).
Model origin: A rule in the code refuses models from developers in countries of concern, wherever they are served. It matches on names (vendor, family, known derivatives), not on provenance, and its list is not complete. No setting lets a refused model through.
AI bill of materials: The models Hexi ships with or names as options are listed in a CycloneDX AI bill of materials, with their developer, their licence and lineage where published, and where requests go. Hexi will accept any other model id you give it, short of that rule. For a copy, write to contact@elovaren.com.

The nervous system

The brain and nerve cord seen from one side, with the first neurons each of seven senses reaches lit in seven colours, and the neurons read back as reactions ringed. — Where Hexi’s senses arrive: the first neurons each of seven senses reaches. Sweet (a success) is green, bitter (an error) red, touch (your message) blue, heat (a busy host) orange, sound (a recording it listens to) purple, fullness (its context filling up) teal and swallowing (what a step spent) yellow. Ringed: the 26 neurons Hexi reads back as its reactions.

What it is: Beside the model, Hexi runs the wiring of a real brain and nerve cord, keeping connections of five or more synapses: 176,422 neurons and the 6.29 million connections between them, simulated neuron by neuron with a published whole-brain model, on a thread of its own. Hexi’s senses drive its sensory neurons, and Hexi reads 26 of its descending, ascending and motor neurons back as reactions: reward, caution, attention and alarm.
What it feels: Each action’s result is a taste: a command or edit that works sweet, any error bitter. Your message is a touch, a busy model host is heat, the pointer moving is wind, the hour drives its clock neurons, and a picture Hexi looks at reaches its photoreceptors.
What it does: It tastes each action’s result and feels each message. In full access mode, where Iris, Hexi’s danger-scoring model, is installed, it can flinch at a command that looks dangerous. Its reflex can only add caution: it can make Hexi ask before a command it would otherwise have run, and it can never let through something the safety rules stopped.
What it is not: It does not make the model smarter. Nothing in its wiring knows language or code; what it adds is a body’s reactions, wired to what happens in a session.
More, where installed: On a computer that has them, a cubic millimetre of real visual cortex (the MICrONS wiring), which sees what Hexi looks at and learns whether pictures like it go well, and a model of whole-brain rhythms (The Virtual Brain), which reads the session as calm, busy or strained; each can move Hexi’s confidence in its work a little, and steers nothing by itself. And eyes that see motion (flyvis, a model of an insect eye’s motion pathway, wired as its connectome is): they watch a second of a video Hexi looks at, or two pictures it looks at moments apart, and pass what moved to the nervous system’s motion neurons, and to its looming detectors when something comes straight at it.
A light brain for small computers: On a computer with less than 6 GB of memory, Hexi runs the wiring of an insect larva’s brain instead, keeping connections of two or more synapses: 2,952 neurons and 300,565 of its 352,611 synapses (Winding, Pedigo et al., Science, 2023), with a body of its own. It tastes, feels touch and warmth, learns in its own learning centre, and flinches at about the same danger score as the full brain.
Credits: The wiring is open research data, used under its licence, and the pictures on this card are drawn from it. Connectome credits

Learning and memory

The top of the brain from the front, with the learning centre’s cells lit in five colours. — The learning centre on its real wiring: 4,064 Kenyon cells in purple, the 271 projection neurons that bring each smell to them in orange, 97 output neurons in blue, and the dopamine neurons that teach them, 316 for reward in green and 16 for punishment in red.

What a command smells like: Each command becomes a smell: its program, its action, its flags and the kinds of file it names. The brain’s learning centre, its mushroom body on its real wiring, turns the smell into a sparse code, so similar commands are recognised as similar and different ones are not.
What teaches it: Dopamine teaches the learning centre from how actions turn out: a success as reward, a failure as punishment. In this version your allow and refuse answers do not reach it.
Lessons: A command that fails is filed as a lesson, with what fixed it once something does. When a similar command fails later, Hexi tells the model what worked last time. The third failure in a row with nothing changed in between is named as a loop, and the model is told to stop retrying and find the cause.
Expectations and surprises: Before an action, Hexi guesses how likely it is to work here, from how actions like it have gone. When a confident guess misses, that is a surprise: remembered, and told to the model, since something has changed.
Sleep: Between sessions, at most every six hours, the memory sleeps: lessons that are the same trouble are folded into one, what was never confirmed fades after thirty days, and troubles that keep coming back are replayed so they are not forgotten.

Feelings

The brain seen from above, with the arousal neurons lit in red and the neurons they reach in orange. — Seen from above, the brain’s arousal neurons in red, which make octopamine, and the 14,730 neurons they reach in the brain in orange (17,226 with the nerve cord’s). When frustration fires them, every neuron they reach answers up to 8% more strongly for a while.

What they are: Hexi keeps a few readings, most between 0 and 1, worked out from what happened in a session: frustration, confidence, urgency, corrections, how full its context is, and what a turn has spent. Each is a measurement, not an emotion.
How they steer it: Each that should change what Hexi does says so once, when it crosses its mark. Frustration makes it change approach or ask. Low confidence makes it say what is not working. Corrections make it re-read what you asked. A full context makes it read and print less.
Arousal: Frustration also fires the brain’s own arousal neurons, the octopamine ones, and for a while every neuron they reach, 17,226 of them, answers up to 8% more strongly.

Safety and privacy

Before a command runs: By default, edits inside the working folder and commands that provably change nothing go ahead, and every other command and every MCP server’s tool asks you. In full access mode only the guardrails ask, and Iris, a model that judges how dangerous a command looks, and the brain’s escape reflex can add a question where Iris is installed. They can only add a question, never take one away.
The go/no-go gate: A model of the basal ganglia, the brain’s circuit for choosing actions, weighs how dangerous a command looks against how likely it is to work here and how commands like it have gone. It learns from dopamine’s prediction error, from how commands turn out. In this version its reading does not stop a command, and /nerves says so.
A safeguard before sending: A message or picture you send that holds what looks like a password, key or card number is stopped before it is sent. What Hexi’s tools read is not checked this way; only files named as secrets are refused.
Hidden characters stripped: Hidden characters are stripped from what Hexi reads, against prompt injection.
A policy that fails closed: An organisation can give Hexi a policy file. What it does not allow, Hexi does not do, and no user can loosen it.
Government profile: One policy key locks Hexi to Amazon Bedrock in your own AWS GovCloud account, over FIPS endpoints, with Claude or NVIDIA Nemotron models. Hexi’s own tools and background traffic are held except the model route; commands run in the shell and MCP servers the policy allows are not. For government
Signals from you: The pointer moving, how fast you type and your corrections are felt as wind, urgency and a bitter taste. The pointer stays on your machine. How fast you type and how often you correct Hexi are told to the model, and a correction goes to a helper model for a lesson unless /settings turns that off. No keystroke is kept: only how many there were, and when.
What leaves the computer: /privacy in Hexi lists what it sends. Each message goes to the model service, and the lookups a turn makes go where they ask: searches to DuckDuckGo, or to Exa when DuckDuckGo refuses. By default Hexi also sends the words you type to a helper model to mend their spelling, and turns worth keeping and your corrections to one for notes and lessons. It checks for a new version once a day at hexinfra.com, installing it in the background on a copy not set up by the Windows installer or a package manager, reads the model service’s list once a week, and fetches parts the first time they are needed, from GitHub, npm, PyPI and Hugging Face. One switch, essential traffic only, holds all of this except each message and its lookups. It is off by default.
What it keeps: What Hexi learns from its commands is kept in its memory on your machine.

Limits

It can be wrong: Hexi’s answers come from the model it runs on, and they can be wrong. Check anything that matters.
Feelings, plainly: Its nervous system is a simulation, and its feelings are measurements of what happened. They steer what it does; they are not a claim that Hexi is conscious.
Evaluations: No independent evaluations of Hexi have been published yet.