MCP

Use the hosted Memex MCP server when your agent runtime supports the Model Context Protocol. Server URL: https://memex.garden/api/mcp Use the same base URL for both OAuth and API-key clients. Do not point clients at /oauth/* directly unless the client explicitly asks for issuer metadata.

Before your first call

Connect with your client’s OAuth flow. No API keys or bearer-token exports are required for the default setup.
Call discover_actions with the intended Memex operation to get the relevant action IDs and input descriptions.
Parse responses using Response shape.
If credits run out, follow Buy credits.

To add usage balance, open https://memex.garden/pricing in your browser.

Available tools

discover_actions
read_action
write_action

Use the execution tool in the discovered card’s name field. Reads use read_action; mutations and external effects use write_action. Both take the chosen actionId and its input object. The server enforces that separation. Read tools declare readOnlyHint: true; write tools declare destructiveHint: true. Additional app-only and handoff integration tools have their own permission annotations. See the Action catalog. This inventory is registered from the shared action catalogue. Tool identity, descriptions, schemas, ordering, and required scopes must be changed there rather than in the MCP server bootstrap. REST clients can use the equivalent POST /actions/discover endpoint. These wrap the Memex operations documented on Available endpoints. REST-only operations are also documented in Available endpoints. For example, use authenticated REST calls to POST /list-auto-tagging-rules and POST /create-auto-tagging-rule when a client needs to read or create auto-tagging rules. For library.search through read_action:

pass responseMode: "lightweight" for discovery cards: each result has transport-only contentEntityId and url, plus a type-specific semantic context object. URLs and IDs are not intended for model context.
pass responseMode: "compact" for the existing LLM-ready source response, or responseMode: "full" for the rich raw payload. responseMode takes precedence over the legacy raw boolean.
omit raw or pass raw: false to get the default compact response array with type, url, createdAt, title?, text, and optional slim media
pass raw: true to get the richer machine-readable payload with results, referencesByResultId, and related entity lists
if the client asks for llm or full, use llm for raw: false and full for raw: true
compact raw: false search defaults reranking on when the query is eligible for rerank
raw raw: true search defaults reranking off unless you set enableRerank: true
contentTypes accepts a list of exact content_entity.type values: web, pdf, youtube, twitter, instagram, tiktok, facebook, linkedin, pinterest, reddit, chatgpt, claude, annotation, image, transcribedMedia, audioRecording, selector, chatThread, twitterProfile, subreddit, youtubeChannel
to search inside private saved views from MCP or Claude, pass viewIds with one or more view IDs
to search specific feeds, pass feedIds using IDs returned by feeds.list_subscribed
to search across all subscribed feeds only, pass feedScope: "all"
omit both feedIds and feedScope to search the full library
pass feedReadStates: ["unread", "read","pending" \| "processing" \| "processed" \| "failed" \| "archived"] when archived feed entries should be included

For saved views:

use library.search with viewIds when an MCP or Claude client needs to search inside a private saved view
use raw: false or omit raw for the llm response option
use raw: true for the full response option
use authenticated REST POST /create-view and POST /list-views to create or list saved views
use authenticated or public-token REST POST /execute-view-search only when the full/raw shape is acceptable or when searching a public shared view token

For feeds.list_subscribed through read_action, use an empty input object. The response returns feeds, the authenticated user’s subscribed RSS and YouTube feed list. For handoffs.list through read_action:

omit status to list pending/unprocessed handoffs, including items that are not ready for webhook delivery
pass status: "processing", "processed", "failed", or "archived" only when the user asks beyond pending handoffs
use referenceContentEntityId to filter to handoffs that reference a specific Memex content entity
use createdAtFrom and createdAtTo for arbitrary ISO timestamp ranges, or day for a single YYYY-MM-DD day
use requestedDestinationText to filter to a target app, agent, or person
the response includes handoffs; these are operational queue records, not content entities
execute the discovered handoff-draining action only after an agent actually processed a pulled handoff

For sharing.create_link through write_action:

pass a saved contentId returned by library.search or library.save_url
set access to view for read-only sharing or collaborate for collaborative access
the response includes shareUrl, publicToken, access, and accessMode

For sharing.list_links:

omit arguments to list all public content sharing links
optionally filter by contentId, contentType, or access
each link includes shareUrl, publicToken, access, accessMode, contentId, and contentType

Choose an auth mode

Claude Code or raw MCP client

Use API-key auth when the MCP client can send custom headers or query parameters directly. Send the API key as either query parameters:

https://memex.garden/api/mcp?apiKey=YOUR_MEMEX_API_KEY
https://memex.garden/api/mcp?userId=YOUR_USER_ID&apiKey=YOUR_MEMEX_API_KEY

Or as headers:

X-Api-Key: YOUR_MEMEX_API_KEY
X-User-Id: YOUR_USER_ID

If your MCP client already has an OAuth access token or Supabase session JWT, it can also send:

Authorization: Bearer YOUR_BEARER_TOKEN

Request shape

List tools

{
    "jsonrpc": "2.0",
    "id": "tools-list-1",
    "method": "tools/list",
    "params": {}
}

Call a tool

{
    "jsonrpc": "2.0",
    "id": "call-1",
    "method": "tools/call",
    "params": {
        "name": "search_content",
        "arguments": {
            "query": "vector databases",
            "contentTypes": ["web", "pdf"],
            "limit": 5
        }
    }
}

Call `library.search` in raw mode

{
    "jsonrpc": "2.0",
    "id": "call-2",
    "method": "tools/call",
    "params": {
        "name": "search_content",
        "arguments": {
            "query": "vector databases",
            "limit": 5,
            "raw": true,
            "includeNotes": true
        }
    }
}

Call `library.search` inside a saved view

Use this shape for the LLM-ready response:

{
    "jsonrpc": "2.0",
    "id": "call-3",
    "method": "tools/call",
    "params": {
        "name": "search_content",
        "arguments": {
            "viewIds": ["view_id"],
            "query": "vector database benchmarks",
            "limit": 10,
            "raw": false
        }
    }
}

Use this shape for the full response:

{
    "jsonrpc": "2.0",
    "id": "call-4",
    "method": "tools/call",
    "params": {
        "name": "search_content",
        "arguments": {
            "viewIds": ["view_id"],
            "query": "vector database benchmarks",
            "limit": 10,
            "raw": true,
            "includeNotes": true
        }
    }
}

Response parsing

For successful Memex MCP calls:

use result.structuredContent as the parsed object
treat result.content[0].text as the serialized text copy of the same payload

If the tool call fails before structuredContent exists, re-check the request against Available endpoints.

Copyable markdown skill: Skill.md
Hosted plugin bundle for Codex: Codex plugin
Claude plugin: Claude plugin
GenSpark MCP setup: GenSpark MCP
GBrain MCP setup: GBrain MCP
GPT Researcher MCP setup: GPT Researcher MCP
OpenClaw wiring notes: OpenClaw plugin
Obsidian desktop integration: Obsidian plugin

General

API reference

Plugins

Before your first call

Available tools

Choose an auth mode

Claude Code or raw MCP client

Request shape

List tools

Call a tool

Call `library.search` in raw mode

Call `library.search` inside a saved view

Response parsing

​Before your first call

​Available tools

​Choose an auth mode

​Claude Code or raw MCP client

​Request shape

​List tools

​Call a tool

​Call library.search in raw mode

​Call library.search inside a saved view

​Response parsing

​Related agent setups

Before your first call

Available tools

Choose an auth mode

Claude Code or raw MCP client

Request shape

List tools

Call a tool

Call `library.search` in raw mode

Call `library.search` inside a saved view

Response parsing

Related agent setups