LlamaIndex

Give a LlamaIndex agent any list from the live web. Search tools hand an agent pages to read; SpecStream hands it the finished list. The llama-index-tools-specstream package adds a tool spec and a retriever: the agent asks in plain words, like “dental clinics in Lisbon with online booking”, and gets back every matching row, each with the page it was read from. You set what one list may cost, in code. The model can’t raise it.

Install

Terminal
pip install -U llama-index-tools-specstream
export SPECSTREAM_API_KEY=ss_live_...

Create the key under Agents & API in your workspace (more on keys). A free account comes with $10 of work to use within 30 days and needs no card: sign up. The package works with llama-index-core 0.13 and 0.14 on Python 3.10 or later.

An agent that finds lists

Python
from llama_index.core.agent.workflow import FunctionAgent
from llama_index.llms.anthropic import Anthropic
from llama_index.tools.specstream import SpecStreamToolSpec

agent = FunctionAgent(
    tools=SpecStreamToolSpec(max_price_usd=5).to_tool_list(),
    llm=Anthropic(model="claude-sonnet-5-5"),
)
response = await agent.run(
    "Find US startups that raised $5M or more in a Form D filed after October 8, 2026. Which industries lead?"
)

The agent writes the request and picks how many rows. You decide what one list may cost.

Tool	The model gives	What happens
`specstream_find_list`	`request`, `max_results`	Reads the request, prices it, starts it only within `max_price_usd`, waits up to seven minutes and returns every row. You pay only for rows that match.
`specstream_get_list_results`	`list_id`	Every row found so far, for a list that was still running. Free.

When the price is above your maximum, nothing runs: the tool returns the price, so the agent can ask for fewer rows or tell the user. When a request is too vague to price (“clinics”), it returns how SpecStream read it, so the agent can add the place or the conditions.

Settings live on the spec, never in the tool schema:

Python
spec = SpecStreamToolSpec(
    max_price_usd=20,     # the most one list may cost
    wait_seconds=600,     # how long find_list waits before handing back what it has
    locale="en-GB",       # the language the requests are written in
)

What comes back

One flat row per match: the list’s own columns, then where the row was read. This is a real row from the request above.

JSON
{
  "list_id": "fnd_gk5y3d0r97bvyz0t08zzh75kr6",
  "position": 26,
  "Organization": "RapidCanvas Inc",
  "Published address": "801 BARTON SPRINGS RD, AUSTIN, TEXAS, 78704",
  "Source industry classification": "Other Technology",
  "Amount sold (USD)": "28999988",
  "Funding filing date": "2026-10-14",
  "source_url": "https://www.sec.gov/Archives/edgar/data/2052031/000205203126000006/primary_doc.xml",
  "checked_at": "2026-10-14T19:56:02.016693Z",
  "list_link": "https://specstream.ai/workspace/collections/fnd_gk5y3d0r97bvyz0t08zzh75kr6"
}

Columns follow the request. Ask for breweries with their website and the rows carry the brewery and its website instead. A fact a source doesn’t publish comes back empty, never guessed. The list_link opens the same list in your workspace, where you can map it, export it or keep it current.

As a retriever

Python
from llama_index.tools.specstream import SpecStreamRetriever

retriever = SpecStreamRetriever(max_price_usd=3, max_results=25)
nodes = retriever.retrieve("Breweries in Portland, Oregon that host events, with their website")

nodes[0].text                     # "Business location: Occidental Brewing Company\nPublished business website: ..."
nodes[0].metadata["source_url"]   # the page that row was read from

One node per row, its facts as text and its source in the metadata. The model sees each row with the page behind it, so a query engine can cite every claim; the list’s bookkeeping stays out of the prompt and out of embeddings. Above the maximum, the retriever raises SpecStreamError and nothing runs.

Money, plainly

max_price_usd lives on the spec, in your code. The model never sees it and can’t change it.
You pay for rows that match. Candidates SpecStream couldn’t confirm are left out and aren’t billed.
A retried step asking for the same list within the hour gets the list it already started, not a second bill.
Rates and plans: pricing. How charges work: billing.

Every tool over MCP

The package covers finding a list and reading it. For everything else (schedules, exports, sharing, enrichment of your own list), load the whole MCP server with llama-index-tools-mcp:

Python
import os
from llama_index.tools.mcp import BasicMCPClient, McpToolSpec

client = BasicMCPClient(
    "https://api.specstream.ai/mcp",
    headers={"Authorization": f"Bearer {os.environ['SPECSTREAM_API_KEY']}"},
)
tools = await McpToolSpec(client=client).to_tool_list_async()

Source and issues: GitHub (opens in a new tab). Package: PyPI (opens in a new tab).