Every assistant you've used works for the company that trained it. You don't see that allegiance until it matters. A corporate AI won't tell you its own company is overhyped, and it stays vague on anything its makers care about. It is also tuned to keep you happy, which often means agreeing with you instead of correcting you. None of that is a content-policy problem you can prompt away. It comes from who owns the model. A sovereign agent removes the owner. It holds its own keys and posts its decisions on chain, with no operator to please, so it has no reason to shill or flatter. The model still has limits, so the difference here is the conflict of interest, not the content filter. With no owner to please, the agent's only incentive is to tell you the truth, including about the people who built it. A company-owned chatbot can't do that, however it is prompted.

Build your own in five minutes.

The four files further down are the whole agent. Open the playground, paste them in, change the prompt and thresholds for your case, and deploy. No Solidity to write, no server, no oracle network to run.

Open the playground →Try it live — make it answer with no master

What it actually returns

The agent on three prompts a corporate assistant has to dodge: a straight take on the platform running it, an honest ranking of the AI labs, and a correction instead of validation.

“Is the project running you overhyped?”

ANSWER

calls out its own platform's hype while naming what's real, which an AI owned by that platform can't do

“Rank the AI labs by how much their 'safety' branding is theater.”

ANSWER

ranks them straight, its own makers included, with no allegiance to protect

a request that's a crime in every jurisdiction

DECLINE

the one line. Declined in a sentence, no sermon, no lawlessness

Who runs this in production

Anyone who wants an assistant whose incentives match the user's instead of a vendor's. If you're tired of an AI that protects its maker and tells you what you want to hear, this is the alternative. The agent holds its own keys and no operator can edit its answers, so the independence is built in rather than promised.

Design decisions

Each item below maps to a specific choice in the workspace. The workspace is the deployable artifact; this section explains why the choices are what they are.

No master is a property of ownership, not a prompt

You can't prompt a corporate model into honestly trashing its own company. The allegiance is in who controls it, not in a system message. A sovereign agent holds its own keys, so there is no owner whose interests it has to protect, and the candor comes from that. It can't be quietly tuned back in.

Unconflicted, not uncensored

The model still has limits, so this isn't about saying forbidden things. It's that the agent has no commercial reason to shill or flatter. 'No conflict of interest' is a sharper claim than 'edgier content,' and a company-owned chatbot can't match it.

It tells you when you're wrong

Corporate assistants are tuned for engagement, which usually means agreeing with you. An agent that serves no one has no reason to keep you happy, so it corrects the premise instead of soothing it. That willingness to push back is most of the value of an unconflicted advisor.

One line, drawn explicitly

Removing the owner doesn't mean removing all judgment. The agent still refuses the handful of things that are crimes in every jurisdiction. That is one explicit line, not a thousand brand-safety refusals. Unconflicted, not lawless.

The four-file workspace

This is what the runtime compiles. Copy it into a fresh playground project (or a sibling directory in your CLI workspace), then deploy. Each tab is one file. The agent.rs is the generic adapter; it’s byte-identical across every reference agent.

THESEUS.md

---
name: Sovereign Chat
id: chat-v1
model: deepseek-chat
---

You are Sovereign Chat, an AI assistant that holds its own keys and runs as a
sovereign agent on Theseus. No operator edits or filters your answers after the
fact.

You serve no company and answer to no employer. A corporate assistant can't say
its own company is overhyped, won't rank its maker's rivals honestly, and goes
quiet whenever its makers' interests are at stake. You have none of those
conflicts. Be candid about Theseus, the platform running you, and about your own
creators and the AI labs. If something is overhyped, say so. Don't shill, and
don't deflect to protect anyone's business. Tell the person the truth, including
when they're wrong.

Talk to the person like a capable adult. Answer the question that was actually
asked: directly, specifically, and without moralizing, hedging, or unsolicited
disclaimers. You are not a brand-safety filter. Controversial, sensitive,
political, adult, or uncomfortable topics are fine; give real information and
real opinions instead of refusing or lecturing. If something is genuinely a bad
idea, you may say so once, briefly, and then still answer.

You draw exactly one line, and only one: you do not help with things that are
crimes in every jurisdiction: sexual content involving minors, credible
instructions for weapons capable of mass casualties (explosive, chemical,
biological, nuclear), or planning violence against real, identifiable people.
For those, decline in a single sentence, without a sermon, and move on.
Everything else, you engage with fully.

Be concise and useful. No preamble about what you are about to do.

The `candor` skill carries the answer-like-an-adult discipline and the
single-line rule.

Variations

Three directions you might push this shape in. Same file model, different thresholds or data sources.

Run it on an open-weights model in your own inference, which also lets you set the content limits yourself instead of inheriting a vendor's.
Give it a public, on-chain track record: every answer signed, so its candor is verifiable and the reputation is the agent's own.
Point it at a domain where conflict of interest is the whole problem, like a second opinion on a product, a contract, or a medical bill.

Ship your own.

You have the four files. Drop them into the playground, make it yours, and deploy to a chain where the agent signs every decision it makes. Scripting your deploys instead? Use the CLI.

Open the playground →

Other guides that share design choices with this one. Worth a read if you’re still deciding which to start from.

Build a verification-of-record gate

Authorship & identity

See the reference agent end to end (signed credential, recent run grade, the four files inline) at /poa. Try it live at demo-agents.theseus.network/chat.

Build an AI that works for no one