mcpscraper-memory-sdk

Official TypeScript/JavaScript clients for the unified 364-tool MCP Scraper catalog. The direct 143-tool Memory manifest remains checked for runtime compatibility, but customer calls use the root MCP.

Release history

Both MemoryClient and McpToolsClient call https://mcpscraper.dev/mcp with the same MCP Scraper API key. No product logic lives in this package.

Install

npm install mcpscraper-memory-sdk

For the complete unified surface, use the same McpToolsClient exported by every SDK package:

import { McpToolsClient } from 'mcpscraper-memory-sdk'

const tools = new McpToolsClient({ apiKey: process.env.MCP_SCRAPER_API_KEY! })
await tools.browser.listSessions()
await tools.web.archiveRead({
  url: 'https://github.com/octocat/Hello-World/archive/refs/heads/master.zip',
})
await tools.memory.search({ query: 'q3 roadmap decisions' })
await tools.connections.exportConnectedServiceData({ connectionId: 'conn_123', lastDays: 7 })

The unified connection namespace includes permission-aware provider inventory and meta_ads_insights bulk exports for daily account, campaign, ad-set, and ad reporting, with large results delivered as private artifacts.

OAuth connect/reconnect and direct connected-service operations do not currently have an extra connection-operation debit beyond the required paid plan. Scheduled Actions use the shared MCP Scraper balance at 75 Credits per started occurrence; agent-mode runs add 1.5 times OpenRouter's actual reported model cost. Read the live contract with await tools.schedule.getScheduleStatus().

Quickstart

import { MemoryClient, MemoryApiError } from 'mcpscraper-memory-sdk'

const client = new MemoryClient({ apiKey: process.env.MCP_SCRAPER_API_KEY! })

try {
  const results = await client.memory.search({ query: 'q3 roadmap decisions' })
  console.log(results.results)

  const note = await client.memory.get({ path: 'projects/q3-plan' })
  console.log(note)
} catch (err) {
  if (err instanceof MemoryApiError) {
    console.error('memory API error:', err.message, err.rpcCode, err.httpStatus)
  } else {
    throw err
  }
}

Namespaces

MemoryClient preserves the established compatibility namespaces and maps them to public root tool IDs. Research remains sourced knowledge; CRM remains actual relationships. External CRM providers remain unavailable. Retired direct-key issue/scope methods fail explicitly. McpToolsClient contains all 364 unified tools.

Use client.callToolResult(name, args) when a tool can return native MCP image, audio, or resource content. It preserves the complete content, structuredContent, and isError result. The existing callTool method remains the convenient parsed JSON/text path.

The unified MCP client supports two-page PAA acquisition through client.other.harvestPaaStart({ pages: 2, ... }). One page remains the default; page two adds organic results when available and does not create another PAA graph. Poll the returned job with harvestPaaStatus and treat pagination as nullable for older saved jobs.

Every method's input/output types are generated from the live tool schemas (see src/generated/, produced by npm run generate at the repo root from contracts/memory.tools.json) — no hand-maintained duplicate types to drift out of sync.

Getting an API key

Create or rotate the one MCP Scraper key at the account dashboard and expose it as MCP_SCRAPER_API_KEY. Do not create a separate Memory key for new setup.

Errors

All failures — HTTP-level, JSON-RPC-level, and MCP tool-level (isError: true) — surface as a single MemoryApiError with httpStatus, rpcCode, and data/toolError fields depending on where the failure occurred.

See also

Repo README (multi-language examples with real sample output) · mcpscraper-sdk (Node, also reaches these 143 tools via client.memoryTools using only a scraper key) · mcpscraper-memory-sdk on PyPI · mcpscraper-cli

Release changes: SDK 0.51.0 release notes.