SpecStream for LlamaIndex

What you get

• Nodes with clean metadataEach match becomes a TextNode that shows the model its source_url and keeps list IDs out of the prompt and your embeddings.
• Settings on the specThe model fills in a request and a row count while max_price_usd, wait_seconds, and locale stay on SpecStreamToolSpec.
• The full server via McpToolSpecMcpToolSpec loads the rest of SpecStream into the agent, including schedules that rerun a request anywhere from hourly to yearly.

Set it up

1. 1Run pip install -U llama-index-tools-specstream on Python 3.10 or later, with llama-index-core 0.13 or 0.14.
2. 2Create a key under Agents & API in SpecStream and set it as SPECSTREAM_API_KEY, or pass it as api_key.
3. 3Give SpecStreamToolSpec(max_price_usd=5).to_tool_list() to a FunctionAgent, or call SpecStreamRetriever.retrieve with a request.

The full setup guide has every step, with what to do if something doesn’t arrive.

• OpenAI APILive web records as an MCP tool in the Responses API
• LangChainAgent tools and a retriever that returns LangChain Documents
• CrewAILive web rows for the agents in your crew
• VercelAI SDK tools in TypeScript, plus a client for route handlers
• HaystackHaystack Documents from the live web, for pipelines and agents
• BotpressComing soonHub actions that put live web rows in a bot's reply