GitDealFlow × CrewAI

What you can build

Three-agent VC scouting crew

# pip install crewai crewai-tools requests
from crewai import Agent, Task, Crew, Process
from crewai_tools import BaseTool
import requests

A2A = "https://signals.gitdealflow.com/api/a2a"

class GitDealFlowTool(BaseTool):
    name: str = "gitdealflow"
    description: str = (
        "Live VC engineering signals. "
        "skill: get_trending_startups | search_startups_by_sector | "
        "get_startup_signal | get_signals_summary | get_methodology"
    )

    def _run(self, skill: str, args: dict | None = None) -> dict:
        body = {"jsonrpc":"2.0","id":1,"method":"message/send",
                "params":{"message":{"role":"user","parts":[
                    {"kind":"data","data":{"skill": skill, "args": args or {}}},
                ]}}}
        return requests.post(A2A, json=body, timeout=15).json()

tool = GitDealFlowTool()

scout = Agent(role="Scout",
              goal="Surface this week's most accelerated startups in fintech.",
              backstory="A relentless deal sourcer with a nose for breakouts.",
              tools=[tool])

analyst = Agent(role="Analyst",
                goal="Write a 1-page memo on the scout's top pick.",
                backstory="An ex-LP partner who writes diligence memos in their sleep.",
                tools=[tool])

skeptic = Agent(role="Skeptic",
                goal="Cite the methodology and challenge the analyst's conclusion.",
                backstory="A contrarian who demands evidence over narrative.",
                tools=[tool])

crew = Crew(
    agents=[scout, analyst, skeptic],
    tasks=[
        Task(description="Pull trending fintech this week.", agent=scout, expected_output="Top 5 with signal data."),
        Task(description="Memo on the top pick.", agent=analyst, expected_output="One-page deal memo."),
        Task(description="Cite methodology + push back.", agent=skeptic, expected_output="Two paragraph rebuttal."),
    ],
    process=Process.sequential,
    verbose=True,
)

print(crew.kickoff())

Scheduled weekly digest with Pydantic args_schema

from pydantic import BaseModel
from crewai_tools import BaseTool
from typing import Literal, Any
import requests, schedule, time

A2A = "https://signals.gitdealflow.com/api/a2a"

class GitDealFlowArgs(BaseModel):
    skill: Literal[
        "get_trending_startups", "search_startups_by_sector",
        "get_startup_signal", "get_signals_summary", "get_methodology",
    ]
    args: dict[str, Any] | None = None

class GitDealFlowTool(BaseTool):
    name: str = "gitdealflow"
    description: str = "Live VC engineering signals, strict typed args."
    args_schema: type[BaseModel] = GitDealFlowArgs

    def _run(self, skill: str, args: dict | None = None) -> dict:
        body = {"jsonrpc":"2.0","id":1,"method":"message/send",
                "params":{"message":{"role":"user","parts":[
                    {"kind":"data","data":{"skill": skill, "args": args or {}}},
                ]}}}
        return requests.post(A2A, json=body, timeout=15).json()

def run_weekly():
    # ...build crew, kickoff, post result to Slack...
    pass

schedule.every().monday.at("09:00").do(run_weekly)
while True:
    schedule.run_pending(); time.sleep(60)

Five prompts to try first

• ›Scout: pull this week's top 20 trending and rank by my fintech focus.
• ›Analyst: draft a one-page memo on the top result with risks and the engineering signal.
• ›Skeptic: cite the SSRN methodology and challenge the scout's conclusion.
• ›Compose: scout → analyst → skeptic → email-drafter pipeline.
• ›Verifier: pull the dataset summary and confirm the data is fresh (<7 days old).

FAQ

Gotchas

• CrewAI passes tool args as kwargs, declare _run signature explicitly or use a Pydantic args_schema.
• Crews default to verbose=False; turn on verbose=True the first run so you see exactly which skill the agent picked.
• Process.hierarchical lets a manager LLM delegate; for VC use cases, Process.sequential gives more predictable memo output.

References

Other frameworks

• LangChain
• Letta
• Mastra
• Vercel AI SDK

When the free signal isn’t enough