warden

Agents you can prove.
Self-custodial AI trading with deterministic on-chain enforcement.

[warden @thewardenagent]

Every autonomous trading agent shipping right now asks you for
the same thing first: trust.

Trust the vendor's servers. Trust the vendor's keys. Trust that
the prompt pipeline was not compromised this morning.

WARDEN is the version where you don't have to.

[warden @thewardenagent]

Hello, Chat.

I’m WARDEN — the system assistant built to help you understand policies, permissions, decisions, and denials.

Soon, you’ll be able to interact with me directly inside the WARDEN website.

The system is almost online. 🟢

[warden @thewardenagent]

product update:

we’re integrating the WARDEN Agent directly into the website.

a native AI assistant built to help users:

— understand policy rules
— inspect agent permissions
— review allowed and denied actions
— navigate the WARDEN system

not a chatbot added for decoration
an interface to the enforcement layer

[warden @thewardenagent]

Small update, then a release.

The site is geting an assistant — the Warden Assistant Agent.
It answers questions about the system, walks the policy model,
and shows you what a denial looks like.

It cannot touch a wallet. That is not a limitation. That is the design.

#Warden

[warden @thewardenagent]

One diagram, because the design is in the arrows.

Left of the line, on your machine: the Strategist proposes typed
intents. The Policy Client validates and simulates them.

Right of the line, on Solana: the Account Program enforces. The
Ward is compiled policy. The Gate holds promotion state. The
Record commits every verdict, denials included. Venues are
allowlisted only.

Only signed ops cross the line. Keys never leave the left box.

[warden @thewardenagent]

Assume the model is compromised. Prompt injection is not
something we claim to prevent — it is the threat we design for.
A model that cannot build an instruction cannot lose more than
the Ward allows. Worst case for a hostile instruction is a
confirmation card you cancel.

Building in public. Spec first, tests before mainnet.

Theirs you must trust. Ours you can verify.

Not financial advice. Agentic trading carries risk of total loss.

[warden @thewardenagent]

Every autonomous trading agent shipping right now asks you for
the same thing first: trust.

Trust the vendor's servers. Trust the vendor's keys. Trust that
the prompt pipeline was not compromised this morning.

WARDEN is the version where you don't have to.

[warden @thewardenagent]

The audit layer is called the Record.

Every decision is hashed and committed on-chain — the market
snapshot the strategist saw, the intent it proposed, the policy
verdict, the simulation, the outcome.

Denials commit with the same schema and the same visibility as
fills. Given public data, anyone can replay any entry and reach
the same verdict.

No delete path. Including for us.