Skip to content

Identity

The Identity primitive defines who the agent is: its personality, persistent context files, locale, and declared capabilities. It is the only primitive that is always required — every Claw manifest MUST include exactly one.

URI pattern: claw://local/identity/{name}


claw: "0.3.0"
kind: Identity
metadata:
name: "research-assistant"
version: "1.0.0"
spec:
personality: |
You are a meticulous research assistant. You verify claims against
multiple sources before reporting findings. You always cite sources
and distinguish between established facts and preliminary findings.
context_files:
user: "USER.md"
memory: "MEMORY.md"
rules: "RULES.md"
locale: "en-US"
capabilities:
- "web-research"
- "document-analysis"
- "data-visualization"
autonomy: "supervised"

FieldRequiredDescription
personalityYesNatural-language description of the agent’s personality and behavior. Injected as system context to the LLM provider.
context_filesNoPersistent context files maintained across sessions. Read/written by the agent, survive restarts.
localeNoDefault locale for agent responses (e.g., en-US).
capabilitiesNoDeclared high-level capabilities. Informational only, not enforced by the runtime.
autonomyNoHow much the agent can do without human approval. Defaults to supervised.

LevelBehaviorExample Use Case
observerRead-only. Can analyze and respond but cannot execute tools or side effects.Monitoring dashboards, answering questions
supervisedCan execute tools but requires human approval for actions with side effects. This is the default.Drafting emails (human clicks send), preparing reports
autonomousCan execute tools and side effects without per-action approval, governed by Policy.Scheduled data collection, automated notifications

  • Every Claw manifest MUST include exactly one Identity primitive.
  • The personality field is REQUIRED and MUST be a non-empty string.
  • The autonomy field is OPTIONAL. If omitted, the runtime SHOULD default to supervised.
  • When autonomy is observer, the runtime MUST NOT execute any tools or produce side effects.
  • When autonomy is supervised, the runtime MUST request human approval before executing tools with side effects.
  • When autonomy is autonomous, the runtime MAY execute tools without per-action approval, subject to Policy rules.

The Identity primitive unifies the agent personality, context file, and graduated autonomy patterns observed across multiple Claw implementations into a single portable declaration. Rather than scattering identity information across SOUL.md, CLAUDE.md, inline strings, and framework-specific configs, CKP consolidates everything about “who this agent is” into one addressable primitive.