Action catalog

Memex actions have one shared contract across the in-product agent, MCP, and API documentation. The contract records an action’s stable ID, compact purpose, required scope, effect level, approval requirement, and runtime bindings. The MCP server derives its complete tool manifest from this catalogue. The stable gateway consists of discover_actions and execute_action: discovery always returns execute_action plus the selected canonical action ID and current input summary. Named compatibility tools may still be generated for older clients, but discovery never routes new integrations through them. Backend runtime adapters provide only dependency-bound execution and response shaping. Do not add per-tool registrations to the MCP server bootstrap. Do not assume a listed capability is available to every connected agent. The server resolves the caller’s scopes before returning actions, and it remains the authorization authority when an action is invoked.

Discover only what is relevant

MCP clients should call discover_actions with the user’s intended outcome before selecting a tool. It returns at most six compact action cards, each containing the tool name, purpose, input summary, effect (read, write, or external), approval requirement, and required scopes.

{
    "jsonrpc": "2.0",
    "id": "discover-1",
    "method": "tools/call",
    "params": {
        "name": "discover_actions",
        "arguments": {
            "request": "Find my airport notes and make a handoff for my travel agent"
        }
    }
}

The returned cards are intentionally brief. Call execute_action with the returned card’s id as actionId and an input object matching its current inputSummary. REST clients can make the equivalent authenticated request to POST https://memex.garden/api/actions/discover with { "request": "..." }. Its cards use the API binding and report the scope required for invocation.

Current actions

Action ID	Purpose	Input	Runtime binding	Effect	Approval
`library.multi_search`	Search the library broadly	2–6 lightweight discovery searches and optional filters.	agent: `multi_search_content`	read	none
`library.search`	Search the library	Optional query and filters, including exact `contentTypes` (`web` for websites), date sort, and limit. Agent runs use `responseMode: "lightweight"` discovery cards.	agent: `search_content` mcp: `search_content` api: `search-content`	read	none
`library.get_saved_content_by_url`	Get saved content by URL	An exact saved URL plus optional personal and shared annotation flags.	mcp: `execute_action`	read	none
`library.inspect_contents`	Inspect selected saved items	Agent: search-result selections plus a focused inspection reason; API: `contentIds[]` and `focus`.	agent: `inspect_contents`	read	none
`library.get_entities`	Read saved items (legacy)	One or more saved content IDs. Agent runs use `library.inspect_contents` instead.	api: `get-content-entities`	read	none
`library.hydrate_references`	Resolve references	Content IDs, URLs, or content entities to resolve.	agent: `hydrate_references`	write	agent: required
`handoffs.list`	List handoffs	Optional status, date, destination, and reference filters.	agent: `list_handoffs` mcp: `list_handoffs` api: `list-handoffs`	read	none
`handoffs.list_destinations`	List handoff destinations	No input.	agent: `list_handoff_destinations` api: `list-handoff-destinations`	read	none
`agent.request_input`	Ask for clarification	A concise question and optional reason.	agent: `request_user_input`	read	none
`handoffs.create`	Create handoff drafts	Exact handoff titles, descriptions, destinations, and references.	agent: `create_handoffs`	write	none
`handoffs.commit_extraction`	Commit extracted handoff threads	An extraction idempotency key and zero or more structured handoff source packets.	agent: `execute_action`	write	none
`feeds.list_subscribed`	List subscribed feeds	No input.	mcp: `list_subscribed_feeds` api: `list-feeds`	read	none
`handoffs.drain`	Mark a handoff processed	Handoff ID and optional processing metadata.	mcp: `drain_handoff` api: `drain-handoff`	write	mcp: none api: none
`library.save_url`	Save a URL	URL with optional title, description, HTML, and tags.	mcp: `save_content_by_url` api: `save-content-by-url`	write	mcp: none api: none
`annotations.create`	Create an annotation	Exactly one content ID or URL plus a text selection, media time range, or both; optional note text and privacy.	agent: `execute_action` mcp: `execute_action` api: `execute-action`	write	mcp: none api: none
`sharing.create_link`	Create a sharing link	Saved content ID and view or collaborate access.	mcp: `create_sharing_link` api: `ensure-shared-library-entry-link`	external	mcp: none api: none
`sharing.list_links`	List sharing links	Optional content, type, and access filters.	mcp: `list_sharing_links` api: `list-shared-library-entry-links`	read	none
`tags.create`	Create a tag	Tag name and optional tag metadata.	agent: `execute_action` mcp: `execute_action` api: `execute-action`	write	none
`tags.search`	List or search tags	Optional tag name query and optional bounded result limit; omit both to list all tags.	agent: `execute_action` mcp: `execute_action` api: `execute-action`	read	none
`tags.update`	Edit a tag	Tag ID and fields to update.	agent: `execute_action` mcp: `execute_action` api: `execute-action`	write	none
`tags.delete`	Delete a tag	Tag ID.	agent: `execute_action` mcp: `execute_action` api: `execute-action`	write	none
`tags.ensure_on_contents`	Ensure a tag on saved content	Agent: `contentResults` from `library.search` and `tagName`; MCP/API: `contentIds[]` and `tagName`.	agent: `execute_action` mcp: `ensure_tag_on_contents` api: `execute-action`	write	none
`views.create`	Create a view	View name, expression, and optional configuration.	api: `create-view`	write	none
`views.update`	Edit a view	View ID and fields to update.	api: `update-view`	write	none
`views.delete`	Delete a view	View ID.	api: `delete-view`	write	none
`views.get`	View a saved view	View ID or public token.	api: `get-view`	read	none
`views.list`	List saved views	No input.	api: `list-views`	read	none
`views.share`	Share a view	View ID and sharing configuration.	api: `share-view`	external	none
`views.delete_share_link`	Delete a view share link	Shared view link ID.	api: `delete-shared-view-link`	external	none

Execution policy

Read actions execute directly when the connection has api:read.
Write and external actions require api:write; their action contract also marks whether the Memex agent must create an in-product approval proposal.
Mutations normally require a readback completion check. Actions marked as self-verifying, such as tags.ensure_on_contents, enforce and return their own postcondition instead.
The catalogue helps an agent choose tools. It never grants permission and does not replace server-side authorization.

The in-product agent uses the same catalogue before each run to register only the small set of typed tools relevant to its workflow and request. This keeps normal chat turns from carrying unrelated tool schemas.

General

API reference

Plugins

Discover only what is relevant

Current actions

Execution policy

​Discover only what is relevant

​Current actions

​Execution policy

Discover only what is relevant

Current actions

Execution policy