MCP tools overview

Hive exposes MCP tools that your AI agent can call during a conversation. You don't invoke these directly — your agent decides when to use them based on your instructions.

Tool summary

Tool	What it does
remember	Store or update a memory
recall	Retrieve a memory by its key
forget	Delete a memory by its key
list_memories	List all memories with a given tag
search_memories	Search memories by semantic similarity
summarize_context	Summarise all memories on a topic
pack_context	Token-budget-aware context pack for agents

Scopes

OAuth tokens may carry memories:read and/or memories:write. Each scope grants the tools listed below; ping is the one tool that needs only a valid Bearer token.

Scope	Grants access to
memories:read	recall, list_memories, list_tags, memory_history, relate_memories, search_memories, summarize_context, pack_context
memories:write	remember, remember_if_absent, forget, forget_all, redact_memory, restore_memory
(none required)	ping — health check, requires only a valid Bearer token

Tokens issued through the standard OAuth flow get both scopes by default.

How your agent uses the tools

You typically don't need to tell your agent which tool to use — just give natural language instructions:

• "Remember that..." → remember
• "What did we decide about...?" → recall or search_memories
• "List everything tagged..." → list_memories
• "Forget the..." → forget
• "Give me a summary of..." → summarize_context
• "Fill my remaining context window with..." → pack_context

Progress notifications

Tools whose expected duration exceeds ~2 seconds emit MCP notifications/progress events at each major stage of work. Supporting clients (Claude Desktop, Claude Code) render these as a progress indicator or streaming status line.

Currently emitted by:

• search_memories — reports at 3 stages (vector search → hydrate → rank)
• summarize_context — reports at 2 stages (retrieve → synthesise)

Clients that don't support progress notifications ignore them — the tool still returns its normal final result. Emission is best-effort: if the transport rejects a notification, the tool continues without raising.