Skip to main content
🟢 Status: STABLE
Applies to: CXP-Core · CXP-Agent · CXP-Robot
Last updated: 2026-02-01
CXP assumes execution can cause harm (security, privacy, physical safety).

​
Typical controls

  • Authorization: who can run which capsule and which tools/robots
  • Approvals: human-in-the-loop for sensitive actions
  • Scope limits: allowed tools, zones, speed/force envelopes
  • Attestation: confirm target identity and integrity (optional)
  • Audit: receipts must be stored and queryable

​
Robot-specific safety

  • speed/acceleration clamps
  • force/torque clamps
  • no-go zones and proximity stops
  • emergency stop integration

​
Agent-specific safety

  • least-privilege tool scopes
  • rate limiting
  • data access constraints
  • redaction policies

​
This page is in progress

This section is actively being drafted for the CXP protocol. Content may change as the spec evolves and community feedback is incorporated.

​
What will be added here

  • Clear definitions and terminology
  • End-to-end examples (capsule → match → plan → execute → receipt)
  • Normative requirements for compliance (where applicable)
  • Security and safety considerations

​
Contribute

  • Propose changes via /community/cxp-guidelines
  • Submit issues or pull requests in the GitHub repository