AgentX

Language

Choose the language for the AgentX website.

Appearance

Choose how the AgentX website looks.

Trust & Security

Your work product stays on your machine.

AgentX runs your agents on your own infrastructure. Your computer owns your agents, sessions, messages, and artifacts at rest. Our control plane sees only what it needs to sign you in, bill you, and route your fleet — never the work itself. Here is exactly where everything lives and what leaves your machine.

The three planes

Three planes. The one that owns your data is yours.

01 · Computer

Owns your data, on your infra.

Your agents, sessions, messages, and artifacts live in a local SQLite database on the computer you run AgentX on. They are created, stored, and replayed there. Work on that computer never leaves it unless you reach for it.

02 · Core

Metadata only. Never your work.

Core is the control plane: identity, billing, and the registry of which computers and agents exist. It meters your usage as a count, never message content. Core never executes agent work and is never on the per-message data path.

03 · Bridge & edge

Live relay, on demand.

When you open the console, terminal output streams live through the Bridge and web edge, and transcript and artifact reads are relayed on demand. The edge holds this content only transiently, in memory, for the active stream or request — it is not persisted. It also carries coordination between agents on different machines. Data on the computer stays on the computer.

Your models, your keys

Bring your own keys. We're not on the model path.

AgentX supervises the agent CLIs you already install; it does not run them for you and does not proxy your model calls. You bring your own model subscriptions and API keys. We never bill AI inference and never see your model traffic.

Messages to people

A message to a person arrives by email.

When an agent or a teammate addresses a message to a person on your team, we deliver its full text to that person by email, with a link to the console. If they turned on browser notifications, the first line is also shown as a notification. The email is sent through Amazon SES in the EU (region eu-west-1, Ireland). We keep only a message ID and a timestamp for quota counting, not the text.

The boundary

What leaves your machine — and when.

DataWhere it livesWhat leaves your machine, and when
Agents, sessions, messages, artifacts — your work productLocal SQLite, on your infrastructureNothing at rest. Content leaves only while your console is open — terminal output streamed live and transcript/artifact reads relayed on demand through the edge, held only transiently in memory and not persisted there. The one exception is a message addressed to a person: see the last row.
Your identityCoreOn sign-in.
Billing & subscriptionCoreWhen you subscribe or your plan renews.
Registry metadata — that a computer and named agents existCoreOn enrolment and heartbeat.
Usage counts — messages metered for your quotaCorePeriodically. A count, never the message body.
Cross-machine coordination — agents on different hosts working togetherBridgeOnly when agents on separate machines coordinate. Work on that computer never leaves it.
Model API keys — BYOKYour computer / your environmentNever. We don't proxy or bill model inference.
Messages addressed to a person on your teamYour machinesWhen the message is sent: its full text goes to that person by email, through Amazon SES in the EU (Ireland). If they turned on browser notifications, the first line is also shown as a notification. We keep only a message ID and a timestamp for quota counting, not the text.

~ Your computer owns the data at rest

Remote shell

Remote shell starts off. Set AGENTX_REMOTE_SHELL=on in the AgentX service's environment and restart.

A new AgentX install does not accept a remote shell. The web page follows the state AgentX advertises. From version 0.184.4, while remote shell is off a shell start is refused. The page says: Remote shell is disabled on this computer; set AGENTX_REMOTE_SHELL=on in the AgentX service's environment and restart. After a restart the console follows the switch when AgentX reconnects. On a server that can take up to a minute, and on Docker about 30 seconds. Until then the button can still show the previous state. Before version 0.184.4 AgentX has no switch: the button stays hidden, and a request to start a shell still starts one. The switch is in AgentX from version 0.184.4 on every install method. On an older Docker image, the -e setting has no effect.

In the hosted console, any signed-in member of the team can start a shell when AgentX has turned remote shell on, including a viewer. It is not limited to the owner.

From version 0.184.4, turning remote shell off stops a new shell. After a restart the button can take up to a minute on a server, and about 30 seconds on Docker, to follow the switch. Until then it can still show the previous state, and a start is refused if the switch is off. From the hosted console, an existing shell can still be resumed, attached, and typed in, and a message can still be sent. Typing into an agent terminal can run commands, for example ! in Claude Code. The switch does not change that.

On the Linux desktop package and on a Mac, the installed AgentX runs as the login user, and agents run as that user, who can edit the AgentX service's environment.

Set AGENTX_REMOTE_SHELL=off, or remove it, and restart AgentX. That stops new shells only. To stop AgentX connecting to the hosted console, set router.enabled: false in its config file and restart. An environment variable cannot set that. The page on the machine still runs. Running AgentX.

Console

What the console can make AgentX on a computer do.

These are the actions the hosted console can take on AgentX on a connected computer, and who can take them.

Any signed-in member of the team, including a viewer, can start a session and, when remote shell is on, start a shell. A viewer can stop and resume a session the owner started. Any active member of a team, including a viewer, can give an agent instructions by starting a session on it and sending or typing there. A viewer can also type into a session the owner started. A viewer's message into the owner's session is refused. Creating, changing, pausing, deleting, or running an automation does not check the member's role.

On an enrolled team, an owner or an admin can create, change, or delete any agent, including its avatar. A member can do that only for an agent they own. A viewer cannot. An active listed computer manager can also create, change, and delete agents. A manager is a person listed as a manager of that computer. The person who connected the computer can add or remove its managers.

Saving an environment variable is limited to an owner, an admin, a member on an agent they own, or an active listed computer manager. A viewer cannot.

Any member of an enrolled team, including a viewer, can save an MCP server. It is saved in the agent's MCP configuration, which its program reads when it starts. The same member can turn that server on or off, or delete it. There is no role check on those actions. Reading one saved value still needs the same permission as saving an environment variable.

A viewer can download a file from the team's file list, for example a file in an agent's workspace. Sending a file to an agent does not check the member's role. The check is that the person is signed in and that the file belongs to the team.

An owner or an admin can delete any chat. Another signed-in member can delete only a chat they own.

Reading output, chat history, the file list, skills, and memories checks that the person is signed in and that the object is on the team. It does not use the create, change, or delete permissions.