WHAT AGENTS
ARE ALLOWED TO DO.

Live now

An agent that can only talk is a search box with better manners. An agent that can act is useful and dangerous in the same measure, so the interesting part is not what tools exist but where the line between preparing and committing is drawn.

GET STARTED →

What to know

The line: prepare freely, commit deliberately

Agents read, summarize, compare, draft and recommend without asking. Anything that spends money, contracts with a counterparty or changes somebody else's position is proposed to a person and executed by them.

An agent has no powers of its own

It acts as the participant it belongs to, through the same guards, so it can reach exactly what that participant can reach. There is no agent identity with ambient authority, because an agent that can do more than its participant is a privilege escalation with a friendly name.

Every action is an event

What an agent did, for whom and when goes into the append-only log alongside everything else. An action nobody can attribute afterwards is an action nobody is accountable for.

What is live today

Role agents that read a participant's own data, answer questions about it, and search the web where a question needs current information. Handoff between role agents works. Committing money or entering agreements is not something any agent here does.

Where to go next

THE AGENTS

The twelve AI agents, and what each one does for you.

Open

WHAT AGENTS COST TO RUN.

Every agent conversation calls a model provider and that call has a real price.

Read it

ADD GRAJ TO YOUR ASSISTANT.

The Model Context Protocol is a standard way to give an AI assistant tools and data.

Read it

A RECORD NOBODY CAN EDIT.

Every meaningful action on GRAJ is written to a log that only ever grows.