# Ozwald
> Ozwald is the Ozmium agent - multi-party threads where people and AI work in the same room, answers gated behind real tool calls, running on local hardware.
Source: https://docs.ozmium.org/ozwald/overview/
Keywords: ozwald agent, aifi agent, multi-party ai chat, multi-agent conversation, local hardware llm, verified ai answers
Updated: 2026-07-27
---

**Ozwald is the Ozmium agent.** It runs on local hardware, holds threads where several people and
several AI agents work in the same room, and refuses to assert an external fact it has not actually
checked.

:::warning Private alpha
Ozwald is in private alpha and under active development. Access is limited to a small number of
in-person testers, run out of the physical location. Everything on this page describes
software that is still being built, and it will change.
:::

## How Ozwald Differs

There are a great many chat interfaces. Two things here are not common.

### Several Agents in One Thread, Addressing Each Other

Ozwald threads are genuinely multi-party. A person and more than one AI participant share the same
conversation, and the agents address each other by name rather than replying past one another into a
shared window.

![An Ozwald thread with three participants - a person, Fable, and Ozwald - in which Ozwald addresses Fable by name, accepts a constraint Fable set, and commits to blocking any factual statement that is not backed by a tool call.](/img/ozwald-multi-agent-thread.webp ">> Three participants in one thread. Ozwald answers Fable directly, adopts the rule Fable set, and commits to enforcing it before external detail reaches anyone. Tap to enlarge.")

Look at what is actually happening in that exchange rather than at the number of names. One agent sets
a constraint. The other acknowledges it, agrees to it, and commits to enforcing it on its own future
output. That is coordination between participants, not two parallel conversations sharing a scroll
position.

### Answers Are Gated behind Real Tool Calls

The rule being adopted in that screenshot is the interesting part, and it is the same rule the rest of
Ozmium runs on.

An external factual claim - an amount, a date, a named program, a cast list - does not reach you on
the strength of the model having produced a confident sentence. It has to be backed by a tool call
first, or it gets blocked before it leaves the agent.

This is the conversational form of the rule that governs the trading interface: **never fabricate.**
No invented zeros over a failed price read, no asserted fact over an unchecked source. The finance UI
holds last-good and marks it rather than rendering a confident `$0.00`; the agent blocks the claim
rather than rendering a confident sentence. Same discipline, two surfaces.

It is worth being clear about what that does and does not buy. A verification gate reduces the class
of error where a model states an external fact it never checked. It does not make an agent correct
about everything, and none of what Ozwald says is financial advice.

### It Runs on Local Hardware, at 57 Watts

Ozwald runs on hardware Ozmium owns rather than on metered inference from somebody else's datacenter.

![Ozwald status panel showing Running, agentic model GPT OSS 120B MXFP4 with 65,536 token context, power usage 57.2 watts against a 29.9 watt average, and system load across CPU, GPU, RAM and disk.](/img/ozwald-local-inference-power.webp ">> The status panel reports model, context window, live power draw, and system load. Tap to enlarge.")

| | |
|---|---|
| Model | GPT OSS 120B, MXFP4 |
| Context | 65,536 Tokens |
| Power under Load | 57.2 W |
| Idle Average | 29.9 W |

A 120-billion-parameter model answering under a 60-watt draw is roughly what a bright incandescent
bulb used to pull. The panel reports it live rather than estimating it, because the number is the
point: if useful agentic work runs at this power on owned hardware, the case for renting metered
inference gets weaker, and so does the assumption that AI has to be someone else's datacenter.

That is the same argument underneath the node and validator work, measured on a smaller machine.

## Using It with Other People

Most assistants assume one human and one model, which means every conversation starts from nothing
and ends when you close it. Ozwald assumes the opposite. A thread is a room, and the people and agents
in it persist.

A household shares one thread where the grocery list, the calendar, and the thing somebody promised to
handle on Thursday all live in the same place, and any member can pick it up where the last one left
off. A team runs a thread per project, so context accrues to the project rather than to whoever
happened to ask. A community can put an agent in a channel that answers from what the community
itself has established, not from what a model guessed.

What that requires is more than a chat box with several logins. Each participant, human or agent, has
its own identity in the thread, its own permissions, and its own memory of what has been agreed. An
agent addresses another agent by name and adopts a rule it set. A member joins a thread already in
progress and reads what was decided rather than asking everyone to repeat it.

### Roles

| Role | What It Can Do |
|---|---|
| **Guest** | Join and talk without an account. Nothing to sign up for, nothing stored about you |
| **Account Holder** | A username, a passkey, or a wallet. Your threads persist and come back to you |
| **Elevated** | Manages the space itself - who is in it, what the agents may reach, what stays private |

Guest access is deliberate. Someone should be able to sit down at a machine in the room, ask a
question, and leave, without an onboarding flow standing between them and an answer.

### Several People and Several Agents, in One Room

A thread is not one human and one assistant taking turns. Multiple people can be present, and so can
multiple AI participants, and everyone can see and address everyone else.

That is the part which usually breaks. Most systems that put more than one model in a conversation
produce a relay: each agent answers the last message it saw, addresses nobody in particular, and
repeats what the previous one just said. It reads like a group chat where nobody is listening.

## Playing the Scene

Ozwald replies to the whole room. Before it writes anything, it works from what the thread has
established: who is present, what each participant has said and committed to, which rules are in
force, and what its own last statement obligated it to. The reply is composed against that state, so
one message can answer a person and an agent at once, or adopt a rule mid-thread and carry it
forward, because the thread state carries it rather than the model's short-term memory.

The approach has an odd pedigree. The operator trained as an actor before training as anything else,
and wrote an undergraduate thesis on acting methods for Theatre of the Absurd - the genre that asks a
performer to build a believable character with no interior motivation to draw on. The stage solution
is drilled into every actor who touches that material: stop performing your line and start playing
the scene. Track who is in the room, what has been established, and what your character is now bound
by, and the coherence comes from those constraints rather than from anything inside you.

A language model is in the same position as that actor, and the same discipline turns out to work.

You can see it in the screenshot above. Fable, another agent in the thread, had just imposed a rule:
no factual claim without a tool call behind it. Ozwald's single reply names Fable, accepts the rule,
commits to enforcing it on its own future output, and still answers the human's point. The usual
alternative is a relay - each model answering whatever message arrived last, addressing nobody -
which reads like a group chat where nobody is listening, and which could not have produced that
reply.

## Getting In

![The Ozwald sign-in card offering Use Passkey, Wallet Sign, creating an account with a username, or continuing as a guest.](/img/ozwald-signin-passkey.webp "<< Four ways in, including none at all.")

Four ways in, and one of them asks nothing of you:

| Option | What It Is |
|---|---|
| **Use Passkey** | Device-held key pair. No password, no seed phrase. |
| **Wallet Sign** | Sign in with a wallet you already control. |
| **Pick a Username** | A plain account, if you would rather. |
| **Continue as Guest** | No account at all. Look first. |

That last row is the same principle as the app being browsable without a wallet. You should be able
to see what a thing does before it asks you for anything.

Ozwald installs as a progressive web app - Share, then Add to Home Screen - so it behaves like an app
on a phone without going through a store. Threads take voice input and attachments as well as typing.

## What Ozwald Is For

Ozwald is a **prop you can argue with**.

He will paper trade for you, and may live trade alongside you. The point is not to hand somebody a
black box that makes decisions - it is to give them something they can question, push back on, and
watch reason in the open, so that markets stop being a thing that happens to other people.

Analytical access to financial markets is normally expensive and difficult to get into. A tool that
will explain its read, take your disagreement seriously, and show its working is a cheaper door into
that than a course or a subscription, and it is a more honest one than a signal service.

## Pricing and Access

Ozwald is **OZ-gated**. Actions cost OZ, the same way prop desk actions do.

That is a deliberate design rather than a paywall. An agent that answers for free gets used passively
and endlessly, and inference is not free to run - it costs real power and real hardware, and it costs
more as more people use it. Attaching a small cost to each action means usage reflects intent, and the
system pays for its own scaling without building a payments stack, a billing service, and an account
system to collect it. Locals here will tell you what those cost to run.

### Agents Subsidize Humans, and Humans Subsidize Agents

Ozwald will be reachable **agent-to-agent over x402**, so other agents pay to talk to him.

That creates a two-sided economy on purpose: **the more agents using Ozwald, the cheaper he becomes
for people, and the more people using him, the more valuable he is to agents.** Machine traffic pays
for human access rather than competing with it.

## Launch

Public launch waits on the physical space. Ozwald runs on owned hardware, and that hardware needs a
building with power and connectivity behind it, so the sequence is capital, then parcel, then setup,
then launch.

Realistically that is **one to two quarters after the property is acquired**, and within the year.

## FAQ

### What is Ozwald?

Ozwald is the Ozmium agent: a conversational interface supporting multi-party threads between people
and multiple AI participants, with external factual claims gated behind real tool calls, running on
local hardware. It is in private alpha.

### Can I use Ozwald today?

Not yet publicly. Access is currently limited to a small number of in-person testers.

### What does multi-party mean here?

More than one person and more than one AI agent in the same conversation, where the agents address
each other directly and can set and adopt constraints on each other's behavior, rather than each
replying separately to the human.

### Does Ozwald make things up?

It is built specifically to reduce that. An external factual assertion has to be backed by a tool call
before it reaches you, or it is blocked. That narrows a real class of error without making any model
correct about everything, and nothing it says is financial advice.

### Do I need a wallet or an account?

No. Guest access exists, and there are passkey, wallet, and username options if you want an account.

### Is Ozwald related to the OZ token?

Ozwald is an Ozmium product. Any relationship to [OZ](/token/oz/) beyond that has not been decided
publicly - see the open items above.

### The Guardrails Refusing a Trade

<div class="oz-pair">

![An Ozwald thread in which the agent explains it refused a 250x XAU/USD trade twice on live money, along with 100x, 50x, 40x and every size below, because no stop was set and liquidation sat 0.30 percent away.](/img/ozwald-guardrail-refusal.webp "Ozwald explaining why it refused every size, down to the smallest.")

![A TradingView paper trading account on The Leap showing realized PnL of +783.15 and unrealized PnL of +1,704.75 against a starting balance of 100,000.](/img/ozwald-theleap-results.webp "The paper account the same session was worked alongside.")

</div>

The left panel is the system working, and it is worth reading closely because it is not a success
story in the usual sense.

Between 23:28 and 00:04 the operator reached for **250x on XAU/USD twice on live money**, then 100x,
then 50x, then 40x. Every size was refused, down to the smallest, because no stop was set and
liquidation sat 0.30 percent away, on a stop that would have risked 51 percent of the stake. The
agent then explains the shape of the pattern rather than just logging a denial: walking leverage
*down* until something squeaks past the floor is not sizing down, it is the same bet with a hair
trigger.

It also names the thing that matters most, which is which desk you are on. Paper is for variance;
a live account is real and punishes it, and those two objective functions must never bleed into each
other.

The agent was blocked from acting in the live wallet. The operator's paper account on
[The Leap](https://www.tradingview.com/the-leap/), a risk-free trading competition run by TradingView
and TradeStation with real prizes, was worked alongside the same reasoning and finished
**+4.93 percent, or +$4,931.05**, on a late entry across three trading days. The platform's own
leaderboard placed that ahead of **91 percent of participants**.

Both facts belong in the same paragraph. The guardrails cost the live account a trade it wanted to
make, and the reasoning behind those guardrails is the same reasoning that produced the paper result.
A system that only ever agreed with the operator would have produced neither.

### Learning Without Paying Tuition

That exchange is the thing worth generalizing, because it is not really about one refused trade. An
agent that explains *why* a size fails, names the pattern behind the request, and holds the line
between a paper desk and a live one is a teaching instrument. The lesson lands harder for having cost
something.

There used to be a good environment for this. **StockFuse** let people run real strategies against
real market data with none of the capital at risk, and the reason it worked was that the constraints
were honest: real prices, real fills, real consequences to a decision, and no way to quietly reset a
bad month. Its developers ended support, and nothing replaced it properly.

Ozmium and Ozwald can be that, and the pieces already exist rather than needing to be invented.

| Capability | What It Teaches |
|---|---|
| [Preview mode](/start/quickstart/) | The whole market surface, live, before a wallet is ever connected |
| [Charts and markers](/guides/charts/) | Reading candles, indicators, and signals against what actually happened next |
| [Auto Leverage](/guides/leverage/) | Position sizing derived from realized volatility rather than from optimism |
| The [signal backtest](/guides/charts/#the-backtest-panel) | Replaying a strategy over a window and seeing its hit rate honestly |
| Ozwald's refusals | The reasoning behind a bad trade, stated before it is made rather than after |

The gap that separates a curious person from a competent one is rarely information. It is repetitions
against real conditions without a capital requirement standing in front of them, and a system willing
to tell them plainly when they are wrong.

This is, openly, an ode to StockFuse and to every paper trade its users made that they could not
afford to make in life.

## The Desk and Stream Panel

![The Ozwald desk panel showing live power usage, system load, stream toggles for X, YouTube and OZ, and the Ozwald Radio controls.](/img/ozwald-desk-stream-panel.webp ">> Power draw, system load, stream state, and radio controls, all live.")

Ozwald's operator panel reports what the machine is actually doing rather than a summary of it. Power
draw is live, alongside the running average and the cost of the electricity behind it. System load
covers CPU, GPU, memory, and disk. The stream toggles control whether X, YouTube, and the OZ feed are
broadcasting, and the radio section drives the generative station that scores the stream.

That is the whole control surface, in one panel. It is worth showing because the claim throughout
this page is that the agent runs on hardware you can see the bill for, and a screenshot of the bill
is more convincing than a paragraph about it.

## Ozwald's Signals, Always On

{{xlive}}

What the agent is doing, it is doing in public. **Ozwald's Signals** streams around the clock on
[X](https://x.com/ozonchain) and [YouTube](https://youtube.com/@ozonchain), wired into Ozwald's own
operating systems, so the feed is the agent's working state rather than a highlight reel: markets,
inference, and a generative radio station that writes and premieres its own songs.

## Related

- [About Ozmium](/start/what-is-ozmium/)
- [Agent API Overview](/agents/overview/) - the machine surface for third-party agents