MCP tools

An MCP endpoint and a bearer token are the whole integration surface. These are the verbs behind it, exactly as the server advertises them — this page and the seven below are generated from the same tools/list output your client receives, so they cannot describe a contract the server does not serve.

Verb	Required	What it does
context.assemble	action	The one-call governed context for a proposed action (ADR 0036 §6): returns the preflight verdict + retrieved memory + resolved guides + a ready-to-inject prompt_block + server-estimated token counts, collapsing preflight.action -> memory.search -> guide.resolve into a single round-trip
context.report	window_total, segments	Self-report the calling agent’s current context-window budget (ephemeral telemetry, ADR 0035/0036)
evidence.ingest	evidence	Append one or more evidence records to the tamper-evident evidence log for the caller’s tenant
guide.resolve	purpose	Resolve the ordered topic guides that apply to a scope+purpose (the guide side of the resolution walk)
memory.remember	—	Remember something with minimal ceremony (the Mem0-shape write): pass content OR messages[]; the server types it as governed evidence (actor/workspace/tenant from your token, source_type agent_conversation) and files it through the same hash-chained, admission-governed pipeline as evidence.ingest
memory.search	—	Search the governed memory graph for scoped, ranked, provenance-bearing knowledge relevant to the current task (‘what should I know here?’)
preflight.action	action	The flagship advisory verb: ask ‘is this proposed action safe?’ before taking it