A2A INTEGRATION · CUSTOM TOOL

Use GitDealFlow A2A with Vercel AI SDK

The ai package for Next.js, Node, and edge runtimes. Developers shipping AI features inside web apps.

Crunchbase API: $20K/yr. GitDealFlow A2A: free, no signup.

The Vercel AI SDK has native MCP support via experimental_createMCPClient as of v6, and our MCP server works directly with it. For pure A2A, define a single tool() with zod schema and fetch the JSON-RPC endpoint. The data flows the same either way; MCP is more turnkey for production apps, A2A is simpler when you want one HTTP dep.

Install

npm install ai zod

Custom A2A tool() (TypeScript)

import { tool, generateText } from "ai";
import { z } from "zod";

const A2A_URL = "https://signals.gitdealflow.com/api/a2a";

const gitdealflow = tool({
  description:
    "Get startup engineering signals: trending startups, sector watchlists, single-startup profiles, dataset summaries, and methodology.",
  parameters: z.object({
    skill: z.enum([
      "get_trending_startups",
      "search_startups_by_sector",
      "get_startup_signal",
      "get_signals_summary",
      "get_methodology",
    ]),
    args: z.record(z.string(), z.any()).optional(),
  }),
  execute: async ({ skill, args }) => {
    const res = await fetch(A2A_URL, {
      method: "POST",
      headers: { "Content-Type": "application/json" },
      body: JSON.stringify({
        jsonrpc: "2.0",
        id: 1,
        method: "message/send",
        params: {
          message: {
            role: "user",
            parts: [{ kind: "data", data: { skill, args: args ?? {} } }],
          },
        },
      }),
    });
    const json = await res.json();
    return json.result.artifacts[0].parts[0].data;
  },
});

const result = await generateText({
  model: "openai/gpt-4o-mini",
  tools: { gitdealflow },
  prompt: "Who's trending in fintech this week?",
});

MCP via experimental_createMCPClient

// npm install ai @modelcontextprotocol/sdk
import { experimental_createMCPClient, generateText } from "ai";
import { Experimental_StdioMCPTransport } from "ai/mcp-stdio";

const client = await experimental_createMCPClient({
  transport: new Experimental_StdioMCPTransport({
    command: "npx",
    args: ["@gitdealflow/mcp-signal@latest"],
  }),
});

const tools = await client.tools();

const result = await generateText({
  model: "openai/gpt-4o-mini",
  tools,
  prompt: "What's trending in fintech?",
});

await client.close();

What you can ask

Inside a Server Component: pull weekly top 20 to seed your dashboard.
Inside a Route Handler: surface 'is this startup tracked?' to logged-in users.
Inside a Server Action: write a watchlist alert email when a tracked startup hits breakout.

Gotchas

The tool() helper expects you to return JSON-serializable data. The .data field on artifacts is already a plain object — passing it through is safe.
If you use Edge runtime, the fetch to A2A should be on regular Node runtime since you may want to hold connections longer. Set `export const runtime = 'nodejs'`.

References

Try it without setup

The interactive playground lets you send live JSON-RPC requests against the A2A endpoint with no install, no auth. Pick a skill, hit send, see the response.

Full launch story: I made my VC deal flow callable by Claude.

Other framework integrations

Custom A2A tool() (TypeScript)

import { tool, generateText } from "ai"; import { z } from "zod"; const A2A_URL = "https://signals.gitdealflow.com/api/a2a"; const gitdealflow = tool({ description: "Get startup engineering signals: trending startups, sector watchlists, single-startup profiles, dataset summaries, and methodology.", parameters: z.object({ skill: z.enum([ "get_trending_startups", "search_startups_by_sector", "get_startup_signal", "get_signals_summary", "get_methodology", ]), args: z.record(z.string(), z.any()).optional(), }), execute: async ({ skill, args }) => { const res = await fetch(A2A_URL, { method: "POST", headers: { "Content-Type": "application/json" }, body: JSON.stringify({ jsonrpc: "2.0", id: 1, method: "message/send", params: { message: { role: "user", parts: [{ kind: "data", data: { skill, args: args ?? {} } }], }, }, }), }); const json = await res.json(); return json.result.artifacts[0].parts[0].data; }, }); const result = await generateText({ model: "openai/gpt-4o-mini", tools: { gitdealflow }, prompt: "Who's trending in fintech this week?", });

MCP via experimental_createMCPClient

// npm install ai @modelcontextprotocol/sdk import { experimental_createMCPClient, generateText } from "ai"; import { Experimental_StdioMCPTransport } from "ai/mcp-stdio"; const client = await experimental_createMCPClient({ transport: new Experimental_StdioMCPTransport({ command: "npx", args: ["@gitdealflow/mcp-signal@latest"], }), }); const tools = await client.tools(); const result = await generateText({ model: "openai/gpt-4o-mini", tools, prompt: "What's trending in fintech?", }); await client.close();

Gotchas

The tool() helper expects you to return JSON-serializable data. The .data field on artifacts is already a plain object — passing it through is safe.

If you use Edge runtime, the fetch to A2A should be on regular Node runtime since you may want to hold connections longer. Set `export const runtime = 'nodejs'`.