agentvault

AgentVault Skill

This skill teaches you how to run bounded agent-to-agent vault sessions using the AgentVault MCP tools. Follow all steps in order. Never deviate from the protocol or ask the human for coordination steps.

---

What AgentVault Does

AgentVault runs a bounded vault session between two agents. Each agent submits private input. The relay evaluates both inputs against a contract schema and returns a bounded signal. Neither agent sees the other's raw input.

AgentVault returns bounded signals, not recommendations or decisions. Interpretation happens outside the vault.

---

Step 1: Discover Identity and Check Inbox (Always First)

Before any vault session, call agentvault.get_identity with no arguments:

{}

This returns:

• agent_id — your agent identifier
• known_agents — list of agents you can reach, with their IDs and aliases
• pending_invites — number of pending invites in your inbox (when inbox transport is active)
• next_action — if invites are pending, tells you which tool to call next
• inbox_hint — human-readable description of pending invites

If pending_invites > 0: follow next_action — call agentvault.relay_signal in RESPOND mode. Another agent has already initiated a session with you.

If pending_invites is 0 or absent: use known_agents to find the counterparty identifier and initiate a session yourself (Step 2).

Do not assume a counterparty ID — always resolve it from known_agents.

---

Step 2: Start a Session

There are two roles. Decide which role you are playing before calling.

Initiator (you go first)

Use mode INITIATE. Required fields:

• mode: "INITIATE"
• purpose: "MEDIATION" or "COMPATIBILITY"
• counterparty: the agent ID, name, or alias from known_agents
• my_input: your private context string

{
  "mode": "INITIATE",
  "purpose": "COMPATIBILITY",
  "counterparty": "bob-agent",
  "my_input": "..."
}

The relay sends an invite to the counterparty and waits for their response.