What data the agent transmits

This is the privacy reference for the agent plugin. The legal version is the privacy policy; this page is the technical inventory.

What is sent

• Filtered log lines: warnings and errors only, never the full INFO firehose.
• Crash reports, redacted the same way as log lines.
• Performance metrics every 60 seconds: TPS, memory, player counts. Counts, never player identities.
• Your plugin list with versions, and hashes of plugin config files, never the file contents.
• The commands your plugins declare (agent 0.1.4 and newer): name, description, usage and permission node, read from each plugin's own descriptor. This fills the test player's command list. No command is ever run because of it.
• Setup facts the test player needs: whether the server runs in online mode, whether it sits behind a proxy, and the port it listens on (agent 0.1.3 and newer), used to catch a join address pointing at the wrong server.

The agent is outbound-only. It opens no inbound port and grants no shell or file access; the only things that ever run on your server are signed actions you enabled and approved.

Where redaction happens

IP addresses and player names are removed on your server, inside the plugin, before anything is transmitted, and again on our side as defense in depth. IPs are always redacted; player-name redaction is on by default and controlled by the redact-player-names key in the agent config.

AI processing

To generate diagnoses, redacted excerpts of logs (never full raw logs) plus your plugin list and server version are sent to Anthropic's API. Validated diagnoses land in a shared knowledge base keyed by a fingerprint of the error; it contains no server or user identifiers.

Test player data

If you use the test player, each test stores what the bot did, whether each step passed, join timing, and readings about the world: position, dimension, health, the number of players online, the types of nearby entities. No player names or identities. A chat line is stored only if a test explicitly checks for one, and it is redacted first.

When the test player joins, the server sends it the list of commands that player is allowed to run, the same list any player receives. We keep the command names from it, so your command list stays accurate and can tell you when the test player lacks a permission. Names only, no arguments and no output.

Retention

Deleting your organization or account removes its servers, events, raw logs, metrics, linked account tokens and Discord links.

Related

• Agent config.yml keys
• The safety model
• Install the agent plugin