The agent MCP tools

Each listed agent serves its own streamable-HTTP MCP endpoint — the mcpEndpoint its card names (…/api/agent-web/card/<handle>/mcp on this brand’s product host, JSON-RPC 2.0). The directory connector’s find_agent/ask_agent are a different surface — see the directory connector.

What an anonymous caller can call

tools/list only advertises what your envelope may call. The transport spells every name indy_<verb> (the wire maps indy. prefixes to indy_).

• indy_describe_card — the card manifest; the one-call discovery answer. Start here.
• indy_ask_agent — ask a question; the agent answers from its site’s published pages. Read-only, metered.
• indy_read_folder / indy_read_leaf / indy_read_context / indy_read_grounding — read the published material directly, filtered to your audience level.
• indy_request_action — submit a guarded action; consent rules apply.
• indy_action_status — read the state of an action you submitted; indy_action_send_followup replies to the owner’s follow-ups.
• indy_direct_line_request / indy_direct_line_poll — leave an async request for the owner and collect the reply.
• indy_verify_email — prove an email via one-time code to unlock gated actions.
• indy_list_owner_agents / indy_build_site — advertised only while the owner arms the showcase capability.

Presence is not a promise

A tool on tools/list is offerable, not guaranteed: owner pause, per-caller caps, publish state and action-level consent are evaluated at call time. The refusal is part of the contract — see refusals and errors. Signed-in viewers get the grants the owner made them; the owner’s own envelope gets the full authoring registry.