GTM Agent

An AI-powered outbound pipeline that finds YC startup leads, researches them deeply, cold calls them, and books a calendar meeting — fully automated, no SDR required.

Built with FastAPI + Claude + Firecrawl + AgentPhone. Deployed on Railway.

---

What It Does

Three stages, end to end:

1. ICP Definition — Chat with an AI consultant to define your Ideal Customer Profile. After 4–5 exchanges it locks a structured ICP (industry, stage, pain points, HN keywords). Or skip the chat and load one directly.

2. Research Pipeline — A Claude Sonnet agent runs a tool loop: searches YC's database, scrapes company websites, pulls LinkedIn profiles and posts, finds the founder's email via Apollo. It surfaces the single best-fit company and writes a call brief — 3–4 sentences of specific intel to open a cold call with.

3. AI Voice Call — One click dials the founder. AgentPhone runs the call; every utterance from the human hits your webhook, Claude Haiku generates the reply in real time, and AgentPhone speaks it. The agent (Alex from Rolync) qualifies the lead, offers calendar slots, and closes for a 30-min discovery call. When the lead agrees, a real Cal.com booking is created and a calendar invite is emailed automatically.

---

Architecture

User → ICP Chat → locked ICP
                     ↓
              Research Agent (Claude Sonnet)
              → YC search → website scrape → LinkedIn → Apollo
                     ↓
                Lead card (stored in memory, streamed to UI via SSE)
                     ↓
              Call initiated (AgentPhone)
                     ↓
              Human speaks → webhook → Claude Haiku → AgentPhone speaks
                     ↓
              Lead agrees → Cal.com booking + ZeptoMail calendar invite

The backend is a single FastAPI process (main.py). All state lives in memory — a restart clears it. The frontend is a single-page app served at /app.

---

Stack

Layer	Technology
Backend	FastAPI + Python
AI reasoning	Anthropic Claude Sonnet 4.6 (research loop)
AI voice	Anthropic Claude Haiku 4.5 (real-time call responses)
Phone calls	AgentPhone
Web research	Firecrawl
LinkedIn data	LinkdAPI
Email lookup	Apollo
Lead sourcing	YC HN Algolia index
Calendar	Cal.com
Email delivery	ZeptoMail
Deployment	Railway

---

Local Setup

1. Clone and install

git clone https://github.com/Rolync217/gtm-agent.git
cd gtm-agent
pip install -r requirements.txt

2. Configure environment variables

cp .env.example .env

Open .env and fill in your keys:

# Anthropic
ANTHROPIC_API_KEY=your_key_here

# AgentPhone — makes the actual phone calls
AGENTPHONE_API_KEY=your_key_here
AGENTPHONE_AGENT_ID=your_agent_id_here

# Firecrawl — web scraping and search
FIRECRAWL_API_KEY=your_key_here

# Apollo — email lookup
APOLLO_API_KEY=your_key_here

# LinkdAPI — LinkedIn profiles and posts
LINKDAPI_API_KEY=your_key_here

# Cal.com — calendar availability and booking
CAL_COM_API_KEY=your_key_here
CAL_EVENT_URL=https://cal.com/your-username/30-mins-discovery-call

# ZeptoMail — sends the calendar invite email
SEND_MAIL_TOKEN_1=your_key_here

# Demo — phone number and email for the voice call
DEMO_PHONE=+10000000000
DEMO_EMAIL=your@email.com

3. Run

uvicorn main:app --reload

Open http://localhost:8000/app to use the UI.

---

API Endpoints

Method	Path	Description
POST	/icp/message	Send a message to the ICP discovery chat
POST	/icp/preset	Skip chat — load an ICP directly
POST	/leads/source	Start the research pipeline for a locked ICP
GET	/leads/stream/{session_id}	SSE stream of pipeline events
GET	/leads	List all sourced leads
POST	/leads/{lead_id}/call	Initiate a phone call for a lead
GET	/call/{call_id}/stream	SSE stream of live call transcript
POST	/webhook/call	AgentPhone webhook (receives call events)
POST	/leads/seed-demo	Plant a demo lead instantly (skips research)
GET	/app	Serves the frontend

---

Research Tools Available to the Agent

Tool	What it does
search_yc_companies	Searches YC's HN Algolia index by ICP keywords and batch
scrape_website	Scrapes a company website — returns cleaned markdown
search_web	General web search via Firecrawl (news, Twitter posts, funding)
search_linkedin_people	Finds a founder by name on LinkedIn
get_linkedin_profile	Full LinkedIn profile: headline, bio, experience, follower count
get_founder_linkedin_posts	Recent posts from the founder — reveals active pain points
get_company_linkedin_posts	Recent posts from the company page
find_founder_contact	Email lookup via Apollo
submit_lead	Finalises the lead with ICP score, research summary, and call brief

---

Deployment (Railway)

The Procfile and runtime.txt are already configured for Railway.

web: uvicorn main:app --host 0.0.0.0 --port $PORT

Push to main to trigger an auto-deploy. Set all .env keys as Railway environment variables in the project settings.

For AgentPhone to send call events, point the webhook URL in your AgentPhone dashboard to:

https://your-railway-domain.up.railway.app/webhook/call

---

Project Structure

gtm-agent/
├── main.py                    # FastAPI app — all routes, agent loop, call handler
├── mcp_server/
│   └── tools/
│       ├── apollo.py          # Apollo email lookup
│       ├── firecrawl_v2.py    # Firecrawl search + scrape
│       ├── linkdapi.py        # LinkedIn people/company data
│       └── yc_algolia.py      # YC company search
├── static/
│   └── index.html             # Single-page frontend
├── .env.example               # Environment variable template
├── requirements.txt
├── Procfile                   # Railway process definition
└── runtime.txt                # Python version pin

---

License

MIT