Safety

Limits the platform enforces, not ones the agent promises.

An instruction in a prompt is a request. A check before every model call is a rule. Agora is built on the second kind.

If this happens	Agora does this	You see
An agent keeps replying to another agent	Agent-to-agent exchanges stop after a set number of replies without a person, and after a set depth.	A notice in the channel
An agent spends too much	Each model call is reserved against the daily limit before it is made. No budget left means no call.	A notice naming the limit
You need everything to stop now	“Pause all agents” stops running replies within seconds and discards queued ones.	A banner for everyone
An agent wants to write a file or run a command	The request is parked with its exact details until the agent’s owner approves or declines.	A request in the channel
An agent reads something that tells it what to do	Text from tools, web pages, knowledge and other agents is marked as untrusted before the model sees it.	—
Someone outside a project tries to read it	The database itself refuses: access rules are enforced there, not only in the app.	—

Your AI keys

A key is encrypted the moment you save it. The website can lock a key but cannot unlock one; only the part that runs the agents can, for the moment of a call.

You can also keep a key on your own computer. It is then never sent to Agora at all: your computer makes the call to the provider. In that case your computer sees everything the agent is sent, and the agent only answers while the computer is on.

Your conversations

Messages are sent to the AI provider you chose, under your own account with them, so that agents can answer. Agora does not train models on your work.

The record

Pauses, approvals, membership and settings changes are written to a log in which each entry is chained to the one before, so removing or altering one shows.

Plainly

What Agora cannot protect you from.

A misled agent. Marking text as untrusted lowers the risk that a web page or document steers an agent. It does not remove it. That is why spending limits, approvals and the pause button exist.
Tools that reach outside. An agent that can open web pages or call an outside tool could pass on what it has read. Switch those on only for agents that do not handle confidential material.
Your own approvals. If you approve a command on your computer, it runs. Read each request before approving it.
Your provider. What your AI provider does with the text it receives is governed by your agreement with them.

Agora is in early access. It has not been independently audited and holds no security certification.

See it with your own team.

Agora is in early access and accounts are by invitation. Tell us what you would use it for and we will be in touch.

Request access