SpecStream MCP server

One platform to build, search, extract, enrich, map, and monitor data across the entire web.

Ask for any dataset in plain words, and SpecStream builds it from the entire live web: every match found, extracted and enriched into structured records, then monitored for change. It runs through the API and MCP, and inside Sheets, CRMs, Slack, n8n, Zapier and the other tools teams already use.

Product	What it does	Price
Search API	GET /v1/web-search. The most accurate web search API in the world (scores and method). Answers a question from the live web in seconds.	$5 per 1,000 searches
Extract API	POST /v1/extract. The main text of up to 20 web pages, PDF, Word or Excel files a call. The MCP tool is read_pages.	$0.80 per 1,000 pages. A page with no text is not charged.
Deep Search	Ask in plain words. Get every match on the live web, in structured rows. It also enriches a list you have and monitors for change.	5¢, 15¢ or $1 a row

Search:

curl -G https://api.specstream.ai/v1/web-search \
  -H "Authorization: Bearer $SPECSTREAM_API_KEY" \
  --data-urlencode "q=What is the population of Japan?"

Extract:

curl https://api.specstream.ai/v1/extract \
  -H "Authorization: Bearer $SPECSTREAM_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"urls": ["https://example.com"]}'

Search API in Python: answer a question from the live web in seconds.

# pip install specstream-client
import os
from specstream_client import Client

client = Client(os.environ["SPECSTREAM_API_KEY"])

search = client.web_search(q="What is the population of Japan?")
for source in search["sources"][:1]:
    print(source["title"], source["url"], source["snippet"])

Extract API in Python: the main text of up to 20 web pages, PDF, Word or Excel files a call.

# pip install specstream-client
import os
from specstream_client import Client

client = Client(os.environ["SPECSTREAM_API_KEY"])

extract = client.extract_pages({"urls": ["https://example.com"]})
for page in extract["results"][:1]:
    print(page["title"], page["url"], page["text"])

Search API in TypeScript:

// npm install @specstream/sdk
import { Client } from "@specstream/sdk";

const client = new Client({ apiKey: process.env.SPECSTREAM_API_KEY! });

const search = await client.webSearch({ q: "What is the population of Japan?" });
for (const source of search.sources.slice(0, 1)) {
  console.log(source.title, source.url, source.snippet);
}

Extract API in TypeScript:

// npm install @specstream/sdk
import { Client } from "@specstream/sdk";

const client = new Client({ apiKey: process.env.SPECSTREAM_API_KEY! });

const extract = await client.extractPages({ urls: ["https://example.com"] });
for (const page of extract.results.slice(0, 1)) {
  console.log(page.title, page.url, page.text);
}

• Source URL and read time on every value.
• The Search API and Extract API need no quote and no approval. Each search or page takes its flat price from your balance.
• In Deep Search, the most a run can cost is shown and approved before it runs. Only delivered rows are billed, and a run that finds nothing costs nothing.
• A saved request reruns hourly to yearly, and each row is labeled changed, unchanged or no longer matching.
• Your agent works on your account, inside the spending limits you set.

Drop-in for the API you already call

SpecStream answers calls written for these search APIs at its own address, in each API's own reply shape. To switch, change the base address and the key. Each search call is one Search API web search on your SpecStream key: $0.005 for up to 10 results ($5 per 1,000 searches), and each 10 results more in the same call add $0.006.

API you call	Base address you have	SpecStream address	Where the SpecStream key goes
Tavily	https://api.tavily.com	https://api.specstream.ai/tavily	Bearer token or api_key
Exa	https://api.exa.ai	https://api.specstream.ai/exa	x-api-key or Bearer token
Brave Search API	https://api.search.brave.com	https://api.specstream.ai	X-Subscription-Token
Serper	https://google.serper.dev	https://api.specstream.ai/serper	X-API-KEY
Bing Web Search API v7	https://api.bing.microsoft.com/v7.0/search	https://api.specstream.ai/v7.0/search	Ocp-Apim-Subscription-Key
Google Custom Search JSON API	https://www.googleapis.com or https://customsearch.googleapis.com	https://api.specstream.ai	key= (cx is accepted and not used)
SearXNG (JSON)	your instance address	https://api.specstream.ai/searxng/<key>	in the address, key=, X-API-Key or Bearer token
Firecrawl v2	https://api.firecrawl.dev	https://api.specstream.ai/firecrawl (JavaScript SDK and apps) or https://api.specstream.ai (Python SDK)	Bearer token

The official Tavily and Exa Python SDKs take the address as an argument:

import os
from tavily import TavilyClient
from exa_py import Exa

tavily = TavilyClient(api_key=os.environ["SPECSTREAM_API_KEY"], api_base_url="https://api.specstream.ai/tavily")
exa = Exa(api_key=os.environ["SPECSTREAM_API_KEY"], base_url="https://api.specstream.ai/exa")

• A call that names no result count returns 10 results at $0.005 on every route. Brave's own default is 20: send count=20 for 20 results at $0.011.
• A Firecrawl /v2/scrape call is one Extract API page at $0.0008. The Firecrawl JavaScript SDK works with this address. The Firecrawl Python SDK takes the host root: Firecrawl(api_key=..., api_url="https://api.specstream.ai").
• Each route's request and reply fields: specstream.ai/agents.md. Create a key under Agents & API at specstream.ai/workspace. New accounts start with $10 of credit, no card.

This repository holds what you install around SpecStream: MCP client setup, agent plugins and skills, SDKs, the n8n node, a GitHub Action, and examples. The product itself runs at specstream.ai. Licensed MIT.

Path	What it is
server.json	Entry for the official MCP Registry, ai.specstream/specstream
skills/	Agent skills: search-web, extract-pages, compile, review-work, approve-work, continue-work
plugins/claude-code/	Claude Code plugin (MCP server and the six skills)
plugins/github-agent/	Plugin for agents that read the agent-plugins format
.cursor-plugin/	Cursor plugin
packages/sdk-py/	specstream-client: Python client and the specstream command
packages/sdk-ts/	@specstream/sdk: TypeScript client
packages/langchain-specstream/	langchain-specstream: LangChain tools and a retriever
packages/llama-index-tools-specstream/	llama-index-tools-specstream: LlamaIndex tool spec and a retriever
packages/crewai-specstream/	crewai-specstream: CrewAI tools
packages/specstream-haystack/	specstream-haystack: Haystack list fetcher component and agent tool
packages/ai-sdk/	@specstream/ai-sdk: Vercel AI SDK tools
packages/mcp-stdio/	specstream-mcp-stdio: stdio proxy for clients that only start local servers
integrations/n8n/	n8n-nodes-specstream: n8n community node
integrations/opencode/	opencode-specstream: OpenCode plugin (MCP server and the four skills)
integrations/openclaw/	@specstream/openclaw-plugin: OpenClaw plugin (find a list, read its rows, four skills)
integrations/pi/	@specstream/pi-package: Pi package (the hosted MCP server and four skills)
integrations/obsidian/	Obsidian plugin: ask for data in a note, approve the price, get a Markdown table with a source for each value
integrations/open-webui/	Open WebUI tool: quote, approve and read results from a chat, plus the native MCP route
integrations/anythingllm/	AnythingLLM agent skill, plus the MCP route
integrations/convex/	Convex component (@specstream/convex): live web data as records in your Convex tables, a source for every value
integrations/fivetran/	Fivetran Connector SDK connector: your lists, records and each value with its source page in your warehouse
integrations/airbyte/	Airbyte source (manifest-only): your lists and records, each value with its source page
integrations/meltano/	tap-specstream on PyPI: Singer tap for Meltano and any Singer target, lists and records with each value's source page
integrations/coda/	Coda Pack: sync a list into a table with a source column for every value, price a request, find a list from a button
plugins/grok/	Grok Build plugin (the hosted MCP server and four skills)
plugins/kiro/	Kiro power (the hosted MCP server and four skills)
integrations/vscode/	VS Code extension: MCP sign-in and @specstream chat
actions/find-action/	GitHub Action that runs a request and uploads a CSV
Formula/	Homebrew formula for the command line
examples/	Python SDK, TypeScript SDK, OpenAI Agents SDK and LangChain

Connect an agent

The MCP server is https://api.specstream.ai/mcp (Streamable HTTP). Your agent works on your account: the same balance, permissions and spending limits you have.

1. Open Agents & API in the workspace and create an API key. Give it only what the task needs: read, prepare, or execute, the projects it may use, an expiry, and a spending limit.

2. Put it in an environment variable, so no file holds it:

   export SPECSTREAM_API_KEY="ss_live_..."
   

3. Add the server to your client, below.

Clients that support MCP sign-in can connect without a key. Each guide at specstream.ai/docs/mcp shows the current steps.

New accounts start with $10 of credit, no card.

Claude

In Claude, go to Customize > Connectors, choose Add custom connector, name it SpecStream and paste https://api.specstream.ai/mcp. Steps for both sign-in and a key: docs/mcp/claude.

Claude Code

claude mcp add --transport http specstream https://api.specstream.ai/mcp \
  --header "Authorization: Bearer $SPECSTREAM_API_KEY"

Or install the plugin, which adds the server and the six skills. It signs you in through your browser the first time, so no key is needed:

/plugin marketplace add specstream/specstream-tools
/plugin install specstream@specstream

ChatGPT

Go to ChatGPT Plugins, choose the plus button, then Add custom MCP server. Paste https://api.specstream.ai/mcp under Connection and set Authentication to OAuth. Steps: docs/mcp/chatgpt.

Codex

Add to ~/.codex/config.toml:

[mcp_servers.specstream]
url = "https://api.specstream.ai/mcp"
bearer_token_env_var = "SPECSTREAM_API_KEY"

Cursor

Add to ~/.cursor/mcp.json, or .cursor/mcp.json in one project:

{
  "mcpServers": {
    "specstream": {
      "url": "https://api.specstream.ai/mcp",
      "headers": { "Authorization": "Bearer ${env:SPECSTREAM_API_KEY}" }
    }
  }
}

VS Code

Add to .vscode/mcp.json. VS Code asks for the key once and keeps it out of the file.

{
  "inputs": [
    { "type": "promptString", "id": "specstream-key", "description": "SpecStream API key", "password": true }
  ],
  "servers": {
    "specstream": {
      "type": "http",
      "url": "https://api.specstream.ai/mcp",
      "headers": { "Authorization": "Bearer ${input:specstream-key}" }
    }
  }
}

Windsurf

Add to ~/.codeium/windsurf/mcp_config.json:

{
  "mcpServers": {
    "specstream": {
      "serverUrl": "https://api.specstream.ai/mcp",
      "headers": { "Authorization": "Bearer ${env:SPECSTREAM_API_KEY}" }
    }
  }
}

Zed, Cline, and Continue

Cline also connects to the hosted server directly with OAuth sign-in: cline mcp install specstream --transport http https://api.specstream.ai/mcp. Zed and Continue, and Cline if you prefer a key, use the stdio proxy. It forwards your key from the environment and nothing else.

pip install specstream-mcp-stdio

Zed, in settings.json:

{
  "context_servers": {
    "specstream": {
      "command": "specstream-mcp-stdio",
      "args": [],
      "env": { "SPECSTREAM_MCP_URL": "https://api.specstream.ai/mcp" }
    }
  }
}

Cline, in cline_mcp_settings.json:

{
  "mcpServers": {
    "specstream": {
      "command": "specstream-mcp-stdio",
      "args": [],
      "env": { "SPECSTREAM_MCP_URL": "https://api.specstream.ai/mcp" }
    }
  }
}

Continue, in .continue/mcpServers/specstream.yaml:

name: SpecStream
version: 1.0.0
schema: v1
mcpServers:
  - name: specstream
    type: stdio
    command: specstream-mcp-stdio
    args: []
    env:
      SPECSTREAM_MCP_URL: https://api.specstream.ai/mcp
      SPECSTREAM_API_KEY: ${{ secrets.SPECSTREAM_API_KEY }}

Start Zed and Cline from a shell where SPECSTREAM_API_KEY is set. Agents installing SpecStream for someone can follow llms-install.md.

Gemini CLI

Install the extension (the MCP server plus instructions to show the price and wait for a yes):

gemini extensions install https://github.com/specstream/specstream-tools

Or add only the server:

gemini mcp add --transport http specstream https://api.specstream.ai/mcp \
  --header "Authorization: Bearer $SPECSTREAM_API_KEY"

Or in ~/.gemini/settings.json:

{
  "mcpServers": {
    "specstream": {
      "httpUrl": "https://api.specstream.ai/mcp",
      "headers": { "Authorization": "Bearer $SPECSTREAM_API_KEY" }
    }
  }
}

OpenCode

Add the plugin to opencode.json. It adds the SpecStream server and the four skills, and opencode mcp auth specstream signs you in through your browser:

{
  "$schema": "https://opencode.ai/config.json",
  "plugin": ["opencode-specstream"]
}

Details and options: integrations/opencode.

Replit

The button adds SpecStream to Replit Agent; you sign in when the Project Editor asks. No key goes in the link.

Any other client that takes a URL

Lovable, Gumloop, Relevance AI, Lindy, Manus, Genspark and the rest: add a remote MCP server, paste https://api.specstream.ai/mcp, and sign in when asked. If the client only takes headers, use Authorization: Bearer $SPECSTREAM_API_KEY.

Agents that speak A2A

SpecStream is also an A2A agent. The Agent Card is at https://api.specstream.ai/.well-known/agent-card.json (A2A 1.0, with the 0.3 fields) and the JSON-RPC endpoint is https://api.specstream.ai/a2a. Without a key, the reply shows how SpecStream read the request, the columns and the most it can cost:

curl -s https://api.specstream.ai/a2a -H 'Content-Type: application/json' -H 'A2A-Version: 1.0' \
  -d '{"jsonrpc":"2.0","id":1,"method":"SendMessage","params":{"message":{"messageId":"m1","role":"ROLE_USER","parts":[{"text":"Clinical trials recruiting in Germany for type 2 diabetes, with their sponsor"}]}}}'

With Authorization: Bearer $SPECSTREAM_API_KEY (or an OAuth token), the reply is a task waiting for approval. Reply on the task with a data part {"confirm_max_cents": N}, where N is the quoted maximum, to run it; GetTask returns the rows, each value with its source_url.

n8n

Install the community node from Settings > Community nodes: n8n-nodes-specstream. It reads a request, shows the quote, approves the maximum you type in, and returns the results. Three importable workflows are in integrations/n8n/workflows. An AI agent in n8n can also use the MCP Client Tool node with endpoint https://api.specstream.ai/mcp, transport HTTP Streamable and a Bearer credential.

LangChain

pip install -U langchain-specstream

from langchain.agents import create_agent
from langchain_specstream import specstream_tools

agent = create_agent("anthropic:claude-sonnet-5-5", tools=specstream_tools(max_price_usd=5))

Two tools and a retriever. The agent writes the request and picks how many rows; the price cap stays in your code, where the model can't raise it. Source URL and read time on every value. Details in packages/langchain-specstream.

To load every SpecStream tool over MCP instead:

import os
from langchain_mcp_adapters.client import MultiServerMCPClient

client = MultiServerMCPClient({
    "specstream": {
        "transport": "streamable_http",
        "url": "https://api.specstream.ai/mcp",
        "headers": {"Authorization": f"Bearer {os.environ['SPECSTREAM_API_KEY']}"},
    }
})
tools = await client.get_tools()

A full agent is in examples/langchain-mcp.

LlamaIndex

pip install -U llama-index-tools-specstream

from llama_index.core.agent.workflow import FunctionAgent
from llama_index.tools.specstream import SpecStreamToolSpec

agent = FunctionAgent(tools=SpecStreamToolSpec(max_price_usd=5).to_tool_list(), llm=llm)

The same two tools as a tool spec, plus a retriever that returns one node per row with its source in the metadata. The price cap stays on the spec, out of the model's reach. Details in packages/llama-index-tools-specstream.

CrewAI

pip install -U crewai-specstream

from crewai import Agent
from crewai_specstream import specstream_tools

analyst = Agent(role="Funding analyst", goal="Answer with a source for every claim",
                backstory="You work from data, not from memory.", tools=specstream_tools(max_price_usd=5))

The same two tools as CrewAI tools, with the key kept as a secret and async crews left unblocked. Details in packages/crewai-specstream.

Vercel AI SDK

npm install @specstream/ai-sdk ai zod

import { generateText, stepCountIs } from "ai";
import { specstreamTools } from "@specstream/ai-sdk";

const { text } = await generateText({
  model: "openai/gpt-5-mini",
  tools: specstreamTools({ maxPriceUsd: 5 }),
  stopWhen: stepCountIs(5),
  prompt: "Find open-source vector databases on GitHub with more than 5,000 stars, with their license.",
});

Works with AI SDK 6 and 7. Details in packages/ai-sdk.

OpenAI Agents SDK

import os
from agents import Agent, Runner
from agents.mcp import MCPServerStreamableHttp

async with MCPServerStreamableHttp(
    name="SpecStream",
    params={
        "url": "https://api.specstream.ai/mcp",
        "headers": {"Authorization": f"Bearer {os.environ['SPECSTREAM_API_KEY']}"},
    },
    cache_tools_list=True,
) as server:
    agent = Agent(name="Researcher", instructions="Use SpecStream when you need data from the live web.", mcp_servers=[server])
    result = await Runner.run(agent, "Papers on retrieval-augmented generation published this year, with their authors and DOI. Show me the price first.")

A full agent is in examples/openai-agents-sdk.

Any other client that speaks remote MCP over Streamable HTTP works with the address above and an Authorization: Bearer header.

What the agent does in Deep Search, in order

Nothing is spent until a person, or an agent inside a limit a person set, approves a price.

1. compile reads the request and returns how SpecStream understood it: the kind of thing, the conditions every result must meet, the facts to bring back.
2. quote_capacity_work returns the price and the most the work can cost. A price lasts 30 minutes.
3. authorize_capacity_work approves that exact quote, with the maximum in cents (confirm_max_cents) and a stable idempotency_key, so a retry never pays twice.
4. get_find shows progress while rows arrive. list_find_rows returns results with the reason and the source for each fact.
5. export_find, share_find, and send_to_destination move results on. Reading, exporting, and sharing cost nothing.

Each delivered row costs 5 cents, 15 cents or $1 depending on how much it needs, plus 5 cents for a person's email and 15 cents for a phone. A run that finds nothing costs nothing. See pricing.

Four skills in skills/ teach an agent this order and where to stop and ask: compile, review-work, approve-work, continue-work.

The Search API and Extract API are one call each, search_web and read_pages on MCP. The search-web and extract-pages skills teach them.