Skip to main content
The hook payloads Cursor, Claude Code and Codex hand to the endpoint agent contain source code, prompts, model output, file contents and tool arguments. By default none of that leaves the machine. The agent applies an explicit allowlist: only the fields below are ever written to the buffer, and the ingest endpoint rejects any event carrying a field not in the schema. Prompts, responses and tool content are sent only when an admin turns on full-content capture.

Stays on the device by default

What is sent

Every event carries: Events also carry these fields when the coding agent supplies them: Once per run, the drain also reports on the device itself:

MCP server identity

To answer “is this MCP server what it claims to be”, the agent reports each configured server’s command and URL rather than only a hash, since a hash cannot be triaged. Before they leave the device, credential-shaped flags and values in the command (--token=..., KEY=value, the value after --header) are replaced with <redacted>, paths under the home folder start at ~, and every query-string value in the URL is stripped, keeping only the scheme, host, path and parameter names. Where it can, the agent adds the package name and version and a SHA-256 of the launch binary.

Configuration snapshot

Periodically the agent reports how each coding agent is configured so misconfigurations can be flagged: the permission mode, the tool names in permission rules (never the paths, domains or commands they name), Codex’s sandbox mode and approval policy, a count of additional directories, each MCP server’s identity as above, the names (never values) of environment variables and headers each MCP server is given, which hook events have handlers and how many, the number of rules files per scope (never their contents or paths), and aggregate project trust levels for Codex (never project paths).

Full-content capture

Full-content capture is off by default. An admin turns it on for the whole organization with Capture full content on the Agent Hub Settings tab. Devices pick up the change on their next policy fetch, and stop sending content if they have not fetched one for 24 hours. While it is on, each event also carries a body with what the call said, ran and got back:
  • Secrets are masked on the device. Before a field is written to the buffer, every match of the gitleaks rule set is replaced with [secret:<rule id>], and the event is labeled secret:<rule id>. This happens whatever the masking setting.
  • Each field is capped at 32 KiB on the device, after masking.
  • Email addresses and phone numbers are masked on ingest. The event is labeled email or phone, and while Mask secrets and personal data before storing is on, which is the default, they are stored as [email:<domain>] and [phone:…<last four digits>].
  • Bodies are stored apart from events, encrypted, and only admins with permission to read agent content can open them.
  • Bodies feed generated text. Cimento titles each session from its first prompt and explains findings from the bodies around them, with a model that runs on Amazon Bedrock in Cimento’s own AWS account.
  • Turning it off stops ingest keeping bodies, and devices stop sending them on their next policy fetch. Bodies already kept are not deleted; they expire on the schedule below.

Retention

Enforcement on both ends

The allowlist is a closed schema. The hook can only serialize the declared fields, and the ingest endpoint validates every event against the same schema with unknown fields rejected. A body that arrives while full-content capture is off is dropped, not stored.

Verify it yourself

The buffer is a readable text file at ~/.cimento/telemetry.ndjson on every device. Open it to see what the next drain will send; the drain adds user_email and the device serial as it sends. The agent has no environment variable or local setting that stops collection. To stop it, click Disconnect on the integration, which uninstalls the agent from the targeted devices.