upAPI

@upapi/mcp

upAPI (upapi.io) is an API marketplace and gateway. One API key and one

pooled quota call every operation in the catalog — social media, search, maps, finance,

developer tools, file and browser operations — over a single uniform HTTP contract, so

integrating the tenth API costs no more than the first.

What it is not: not a workflow or orchestration engine, not an LLM provider, and not a

local scraping library. Each call is one hosted operation on api.upapi.io; chaining calls

is up to your code or your agent.

Canonical URL: https://upapi.io · Docs: https://upapi.io/docs · Agent index:

https://upapi.io/llms.txt

Every public upAPI operation as an MCP tool — web search, SERP, social profiles, dev-tool lookups, geo/finance data — so an agent can call them directly.

There are two ways to connect, and they differ only in how a call is authenticated:

Both expose the same tools with the same schemas, and both bill the same quota.

Hosted — no install

Point any MCP client that supports remote servers at:

https://app.upapi.io/api/mcp

It will walk you through signing in to upAPI in a browser; there is no key to copy. With Claude Code:

claude mcp add --transport http upapi https://app.upapi.io/api/mcp

Local — API key

Create a key by asking your agent to run create_api_key over the hosted connection, or on the API keys page, then:

claude mcp add upapi -e UPAPI_API_KEY=upapi_xxx -- npx -y @upapi/mcp

Claude Desktop (claude_desktop_config.json), Cursor, and Windsurf take the same thing as JSON:

{

"mcpServers": { "upapi": { "command": "npx", "args": ["-y", "@upapi/mcp"], "env": { "UPAPI_API_KEY": "upapi_xxx" } } } }

The key is never validated locally — only checked for presence, so a missing one fails immediately with a readable message instead of surfacing later as an unexplained 401 inside a tool call. Whether a key is real, expired, or over quota is answered at the gateway, the single place that answers it for every machine caller.

Tools

The hosted endpoint serves a compact table by default: two meta-tools plus a few always-on operations, a few kilobytes in total.

A tool table is re-sent as context on every turn, so one tool per operation means tens of kilobytes of JSON Schema per turn and a table large enough to measurably degrade tool selection. search_ops + call_op stays flat as the catalog grows. web-search.post, github-repo.get, and wikipedia-article.get stay on the table as full tools so the common case needs no discovery round-trip.

There are two other tables. ?tools=full gives every operation its own tool:

claude mcp add --transport http upapi 'https://app.upapi.io/api/mcp?tools=full'

?tools=directory gives a curated set of 11 named tools for the flagship operations — page-to-Markdown, screenshot, HTML-to-PDF, PDF text, OCR, transcription, GitHub repo, npm package, Wikipedia, currency — with read tools and write tools listed separately and no call_op. That is the shape AI-directory review criteria ask for (a catch-all dispatcher with a target parameter is a rejection, and so is a listing that advertises scraped sources, which is why the Maps trio and web search stay callable but are not listed there), and it is what the Claude Desktop Extension ships with. ?tools=claude is the same shape over the 8 operations upAPI computes itself (page-to-Markdown, screenshot, HTML-to-PDF, PDF text, OCR, transcription and its result, text analysis) — the Claude connector listing. The local stdio server takes the same names in UPAPI_TOOL_MODE, defaulting to compact — the same few-kilobyte shape the hosted endpoint defaults to. full remains available (UPAPI_TOOL_MODE=full) as the rollback for a client already configured against the one-tool-per-operation shape this server used to default to.

Every mode reaches exactly the same operations — the mode changes what is advertised, never what is allowed. Every tool in every mode carries readOnlyHint, destructiveHint, idempotentHint and openWorldHint, derived from what the worker does rather than from the slug's verb suffix. Operations are named after their slug with . and - replaced by _ (web-search.post → web_search_post), and each advertises the operation's real JSON Schema (formats, bounds, defaults, nullability), because that schema is generated from the worker's own model and passed through untouched.

The hosted server reaches the whole public catalog in every mode; since 2026-09-26 it withholds no category. The local (stdio) server and the Desktop Extension catch up in the next release: the published 0.2.0 bundles 63 operations and still withholds Social Media and Utility in directory mode.

directory mode only changes what is ADVERTISED — the curated flagship set a public listing (like the Claude Desktop Extension) shows.

Descriptions carry the quota cost, so an agent can budget:

Search the web… upAPI operation web-search.post (Search). Costs 25 units of monthly quota

per call.

A failed operation comes back as a normal tool result with isError: true and text leading with upAPI's public error code — RATE_LIMITED, INVALID_INPUT, UPSTREAM_UNREACHABLE. A rate limit also states the wait in seconds. Nothing about a failing operation breaks the session.

No outputSchema is declared, deliberately: MCP requires a server that declares one to return matching structuredContent, and these outputs describe live third-party payloads. One unexpected null would turn a successful call into a protocol error.

Use it from Mastra

The tools work in a Mastra agent directly, without an MCP transport in between. @mastra/core and @mastra/mcp are optional peer dependencies — install them yourself, and import the bindings from the /mastra subpath. Nothing else in this package touches Mastra, which is what keeps a plain npm i @upapi/mcp (and the desktop-extension bundle built from it) small.

import { Agent } from '@mastra/core/agent';

import { createGatewayCaller } from '@upapi/mcp'; import { createUpapiTools } from '@upapi/mcp/mastra';

const agent = new Agent({ name: 'researcher', instructions: 'Research topics using upAPI.', model: /* … */, tools: createUpapiTools({ caller: createGatewayCaller({ apiKey: process.env.UPAPI_API_KEY! }), }), });

Narrow the table with filter when an agent should only see part of the catalog:

createUpapiTools({

caller, filter: (op) => op.category === 'Search', });

createUpapiMcpServer — Code Mode by default

createUpapiMcpServer builds a Mastra MCPServer (@mastra/mcp, also an optional peer). Its default tool surface is Code Mode: search_tools + execute_typescript instead of one tool per operation — measured on the full 99-operation catalog, tools/list drops from 111,016 bytes to 1,069 bytes (99 tools → 2), because the table no longer grows with the catalog at all.

import { createGatewayCaller } from '@upapi/mcp';

import { createUpapiMcpServer } from '@upapi/mcp/mastra';

const server = createUpapiMcpServer({ caller: createGatewayCaller({ apiKey: process.env.UPAPI_API_KEY! }), });

await server.startStdio();

• `search_tools({ query? })` finds the operations this connection can reach and returns them as declare function external_ (...) TypeScript signatures.
• `execute_typescript({ code })` runs a short TypeScript program in an isolated QuickJS sandbox (@mastra/quickjs, another optional peer — no native binary, no filesystem/network/process access beyond the injected external_* functions). Batch several operations into one round trip with Promise.all instead of one tool call each. Killed after 30 seconds. A program that computes for more than 1.5 seconds without awaiting an external_* call, or for more than 5 seconds in total across all such stretches, is stopped with CPU_BUDGET_EXCEEDED: ..., because the sandbox shares the Node event loop with every other request the process serves. Needs quickjs-emscripten 0.31.0 (the exact version @mastra/quickjs 0.1.1 resolves) installed next to it. Both are loaded on the first execute_typescript call, so a server that never runs one (or uses the full surface) does not need them installed.

Every external_* call resolves to the exact same createUpapiTools(options) tool object the full surface (below) calls — there is no second auth path. A denied or failed operation call surfaces inside the sandbox as a thrown Error(" : "), so the guest program's own try/catch can inspect and branch on the code; arguments that fail an operation's input schema throw INVALID_INPUT: the same way. That is separate from a malformed execute_typescript input, which fails the call itself before the sandbox ever starts.

Set MCP_TOOL_SURFACE (or pass surface) to change what is advertised:

MCP_TOOL_SURFACE=full node my-server.js   # one tool per operation, the pre-Code Mode shape

createUpapiMcpServer({ caller, surface: 'full' }); // same thing, set in code instead of env

instructions on the server carries the Code Mode contract automatically whenever a codemode tool is on the table (codemode and both) — a model was not told to write TypeScript against search_tools/execute_typescript otherwise.

Build your own