Agents

An agent is a registered identity in your org — a bot, service, or runtime that needs scoped, audited access to secrets and optional on-chain signing or LLM proxying.

Agents do not get blanket vault access. Humans attach policies that grant specific path patterns; JWT scopes are derived from those policies when agents.scopes is empty.

Lifecycle

1. Register — Human creates agent via dashboard, API, or self-enrollment
2. Policy — Human grants read/write on secret paths (golden path)
3. Authenticate — Agent exchanges ocv_ API key for short-lived JWT
4. Operate — Fetch secrets, sign transactions, route LLM traffic, run automations
5. Offboard — Revoke policies, deactivate agent, rotate keys (revoking access)

Capabilities (per-agent toggles)

Feature	Description	Docs
Secret access	JIT fetch via Agent API or MCP	Agent API
Shroud	LLM proxy with redaction and threat detection	Shroud
Intents	Sign transactions without raw private keys	Intents
Execution Intents	HTTP/GraphQL/DB via credential bindings	Guardrails & Execution
Memory	Scratch, durable, and semantic agent memory	Memory
Channels	Telegram, WhatsApp, Discord messaging	Communication
Delegation	Inter-agent task delegation (human-approved)	Delegation
OIDC federation	Exchange agent JWT for RS256 tokens (WIF)	OIDC federation

Child agents

A parent agent can be given cheap sub-agents for fan-out work — one child per document to summarise, one per lead to qualify — without each one being a full, human-registered agent.

• Created by a human with POST /v1/agents/{agent_id}/children (agents.createChild in the SDK, create_child in Python). Up to 50 per parent; a child cannot have children.
• Bounded by the parent. A child's vault_ids and scopes default to the parent's and must be a subset of them — a superset is refused. Vault policies and guardrails are inherited from the parent (policy lookups include the parent's), so a child can never do more than its parent.
• Its own identity. Each child has its own ocv_ API key, its own memory namespaces (default child:{child_id}) and its own action_approval_policy, so its actions are attributable and its approvals are separate.
• Free of the agent cap. Children do not count against the plan's agent limit. Agent responses carry agent_type (standard | child) and parent_agent_id; GET /v1/agents/{agent_id}/children lists a parent's children.

1claw agent create-child $PARENT summariser-7 --scopes secrets:read --namespaces child:doc-7
1claw agent children $PARENT

curl -X POST https://api.1claw.co/v1/agents/$PARENT/children \
  -H "Authorization: Bearer $USER_JWT" -H "Content-Type: application/json" \
  -d '{"name":"summariser-7","scopes":["secrets:read"],"memory_namespace_allowlist":["child:doc-7"]}'

Signing keys

Humans provision per-chain signing keys (Ethereum, Bitcoin, Solana, XRP, Cardano, Tron). Private keys live in __agent-keys; agents sign via Intents API only.

See Multi-chain signing.

Next steps

• Register an agent
• Enable Intents API
• Fleet management