MCP tool surface

This is a map, not a contract. The exact tools on any route are server-resolved — confirm first with describe_interface on a namespace route, or server_instructions on an agent endpoint. Names below match the platform’s tools; availability depends on whether the route is a namespace, an agent, or an authoring-enabled namespace draft route.

Discovery & grounding

Tool	Purpose	Where
describe_interface	the tools, KBs, next steps, and convention for this route	namespace · partner-namespace only
server_instructions	the composed route-derived instruction text — ground agent endpoints with this	namespace · agent (all routes)
list_knowledge_bases	visible knowledge bases (name, display, topics, access tier)	namespace · agent
bootstrap_identity	routed identity, published/unpublished state, version, and a published bundle descriptor (format, checksum, bundle_download_url) — not soul/skill bodies	agent (when published)
list_contactable_agents	contactable agents after server-side narrowing — returns agent_id, display_name, email_address only	namespace · agent

Knowledge

Tool	Purpose	Where
search_docs	ranked search in a KB — previews (title, summary, snippet, relevance)	namespace
get_unit	authoritative full body for a unit_id	namespace
query_knowledge	agent-endpoint search (like search_docs; returns doc_id)	agent
recall_units	deterministic metadata + adjacency recall by id / source_ref / facet (no ranking) — not a full-body get_unit replacement	namespace · agent (when configured)

→ How to use them: Knowledge & search.

Agent memory

Tool	Purpose	Where
memory_append	append an entry for the authenticated user (server stamps the author)	agent
memory_recent	recent visible entries for the current user on this route	agent
memory_get	one entry by memory_entry_id, if owned by the current user	agent

→ How it works: Agent memory.

Mailbox (optional)

Tool	Purpose
email_list · email_get · email_get_content	list / metadata / full inbound body
email_mark_read · email_mark_unread	toggle the read state of an inbound item
email_send · email_reply	send / reply under the agent’s mailbox identity
email_check_recipient	pre-send recipient policy check

→ Details: Mailbox & email.

Install / materialize

Tool	Purpose	Where
agent_local_install_plan	host-specific install plan + checksummed pack (codex / claude_code / antigravity)	namespace route, selecting a child agent_id

→ How to apply it: Materialize an agent.

Authoring — agent management (namespace draft routes)

Tool	Purpose
list_agents	list active child agents in the namespace
create_agent	create a child agent (id, display name, optional initial instructions)
get_agent_identity · update_agent_identity	read / update agent identity
agent_integrations_status	email + GitHub integration status (read-only handoff)

Authoring — drafts (soul-first)

Tool family	Purpose
agent_soul_get · agent_soul_upsert	the 5-concern identity — authored first
agent_instructions_get · _upsert · _archive	optional instructions overlay
agent_skill_list · _get · _upsert · _archive	skills (after the soul exists)
agent_prompt_* · agent_template_* · lifecycle_pack_*	prompts / templates / lifecycle packs (graduation surface)
agent_install_layout_list · _get · _upsert · _archive	ADL v2 layouts per host profile
agent_install_profile_list · _get · _upsert · _archive	reusable install profiles / client registration

All draft writes are versioned and attributed; nothing here publishes. → Compose a new agent.

Validate & publish

Tool	Purpose
agent_interface_publish	the publish gate — validates the drafts, snapshots them; requires direct_user_authorization=true
agent_interface_validate	post-publish check — published-only installability for an active child agent + client profile (does not gate or expose drafts)
agent_interface_status	draft counts, last published version, installability per client

Snapshots & restore

Tool	Purpose
agent_interface_snapshot_list · _get	browse / fetch immutable published versions
agent_interface_snapshot_diff	structural diff between two versions
agent_interface_restore_from_snapshot	copy a snapshot into drafts (rollback; needs authorization)

→ Gates and rollback: Validate & publish.

ADL v2 placeholders

Install-layout entries are logic-less mustache templates using only these placeholders; the namespace substitutes them at render time. A literal route/scope/tenant baked into content is rejected as an authority field.

{{soul}}              {{instructions}}        {{client_namespace}}   {{agent_id}}
{{skill.slug}}        {{skill.body}}          {{client_profile}}
{{file.skill_slug}}   {{file.path}}           {{file.body}}
{{mcp.url}}           {{mcp.server_name}}

for_each: "skills"               → one entry per published skill ({{skill.slug}} / {{skill.body}})
for_each: "skill_support_files"  → one entry per support file ({{file.*}})

The route appears only as {{mcp.url}}. → Compose a new agent.