Inngest AgentKit

Give an AgentKit agent SpecStream's tools over Streamable HTTP. It can search the live web, read pages and build datasets of structured rows, and inside an Inngest function each tool call runs as its own step.

SpecStream for Inngest AgentKitAgent frameworks & builders

What you need

Node.js, @inngest/agent-kit (this guide follows 0.13.2) and its peers inngest and zod.
A SpecStream API key with a spending limit. New accounts get $10 of credit on sign-up, no card.

1. Create an API key

In your SpecStream workspace, open Agents & API, choose API keys, then Create API key. Give it only the spending limit the agent needs, and keep it in an environment variable where the agent runs:

Terminal
npm install @inngest/agent-kit inngest zod
export SPECSTREAM_API_KEY="YOUR_SPECSTREAM_API_KEY"

2. Add SpecStream to the agent

Put SpecStream in the agent's mcpServers with the streamable-http transport. AgentKit passes requestInit to the MCP SDK's StreamableHTTPClientTransport, so the key goes as a header on every request:

research.ts
import { anthropic, createAgent, createNetwork } from "@inngest/agent-kit";

const researcher = createAgent({
  name: "researcher",
  description: "Answers questions from the live web.",
  system: "Use the SpecStream tools to search the web and read pages. Give the source URL for each fact.",
  mcpServers: [
    {
      name: "specstream",
      transport: {
        type: "streamable-http",
        url: "https://api.specstream.ai/mcp",
        requestInit: {
          headers: { Authorization: `Bearer ${process.env.SPECSTREAM_API_KEY}` },
        },
      },
    },
  ],
});

export const research = createNetwork({
  name: "research",
  agents: [researcher],
  defaultModel: anthropic({
    model: "claude-sonnet-4-5",
    defaultParameters: { max_tokens: 4096 },
  }),
  maxIter: 6,
});

3. Run the network

On its first run the agent lists the server's tools and names each one after the server: specstream-search_web, specstream-read_pages and the rest. The model picks them like any tool you made with createTool.

run.ts
import { research } from "./research";

const run = await research.run("Which Postgres extensions for vector search shipped a release this month?");
console.log(run.state.results.at(-1)?.output);

4
4. Run it inside an Inngest function
Serve the network from an Inngest function as you would any AgentKit network. There AgentKit wraps each MCP call in step.run, named after the tool, so a retry of the function reuses a search that already returned and does not pay for it twice.

What the agent can call

AgentKit adds every tool the server lists. These are the ones a research agent uses most:

specstream-search_web answers a question from the live web, with the results it read and a source URL and read time on each value.
specstream-read_pages returns the main text of up to 20 web pages or documents (PDF, Word, Excel) in one call.
specstream-compile reads a request for a dataset into conditions and columns, and specstream-quote_capacity_work prices it. Neither spends anything.
specstream-authorize_capacity_work starts the run, only with the approved maximum in confirm_max_cents. specstream-get_find and specstream-list_find_rows follow it and read the rows.

Spending limits

Searches and page reads are charged per call with no approval. A dataset run needs a quote, and starts only with that quote's maximum. The key's own spending limit holds over both: a call past it is refused with 403. In a network that should not start paid runs on its own, give the key a low limit, or route to a person before authorize_capacity_work.

OAuth for your users

When each user of your app should spend from their own SpecStream account, pass an MCP SDK OAuthClientProvider as authProvider in the same transport, in place of the header. AgentKit hands it to the transport unchanged.

One search tool without MCP

To give an agent a single search tool with no MCP connection, call the Search API from a createTool handler: GET https://api.specstream.ai/v1/web-search?q= with Authorization: Bearer and your key. The Search API docs list the reply's fields.

Cost

Connecting is free. Each search is one Search API call at $5 per 1,000 searches of up to 10 results, and each page read with text is $0.80 per 1,000 pages. A dataset costs 5¢, 15¢ or $1 a row, plus 5¢ for a person's email and 15¢ for a phone, never more than the approved maximum, and a run that finds nothing costs nothing. How quotes and billing work.

Remove

Take the specstream entry out of mcpServers. Revoke the key under Agents & API to stop it everywhere. Results already delivered stay in your workspace. Other ways to build with SpecStream are in agent frameworks. Questions go to [email protected].

SpecStream for Inngest AgentKit covers what it does in Inngest AgentKit on one page. Every other tool has its guide in Connect your tools. You see the most a request can cost before it starts, and the matches are paid from your balance. Billing has the details. For help, email [email protected].