🤖 Autonomous Sales System (Open-Source AI SDR)

A production-grade, autonomous GTM agent powered by a Model Context Protocol (MCP) server and 30+ specialized business agents.

Replace your $50k/year Sales Development Representative (SDR) with an automated, self-improving system that costs $50/month.

🚀 Why Use This?

Traditional B2B outbound is dead. This repository gives you a production-ready Model Context Protocol (MCP) powered autonomous agent that handles the entire sales funnel. Rather than rigid API hardcoding, Claude AI acts as the brain—seamlessly discovering and routing tasks across 30+ specialized business agents (from scraping Apollo, to analyzing GA4, to auto-replying via Gmail).

🏗 System Architecture

graph TB
    %% Core System Components
    Claude(("🧠 Claude AI<br/>(Reasoning Engine)"))
    
    %% MCP Server Layer
    subgraph MCPServer ["🔌 MCP Server Layer (Tools Registry)"]
        direction TB
        CoreAPI["Core APIs<br/>(Gmail, WhatsApp, Apollo)"]
        GrowthAPI["Growth APIs<br/>(GA4, Google Ads, Iterable)"]
        InfraAPI["Infra APIs<br/>(Webhooks, Scheduling)"]
    end
    
    %% Agent Registry
    subgraph Agents ["🤖 30+ Business Agents Registry"]
        direction LR
        Outbound["Outbound Engine<br/>(Sourcing, Drafting, Booking)"]
        Inbound["Inbound Engine<br/>(Triage, Intent, Routing)"]
    end
    
    %% Durable execution & Storage
    Trigger["⚙️ Trigger.dev<br/>(Durable Background Jobs)"]
    Supabase[("🗄️ Supabase<br/>(Lead DB & Auth)")]
    
    %% Connections
    Claude <--> |"Discovers & Calls Tools"| MCPServer
    Claude --> |"Delegates Tasks"| Agents
    Trigger --> |"Orchestrates Multi-step"| Claude
    Agents --> Supabase

MCP Server Upgrade

This repository now includes a production-style MCP server in mcp_server/. MCP, or Model Context Protocol, is the tool layer that lets Claude discover and call business capabilities at runtime. Instead of hardcoding Gmail, WhatsApp, Google Sheets, Ads, Analytics, Contentful, Iterable, and Trigger.dev orchestration into every agent prompt, Claude gets a stable catalog of tools with clear inputs, validation, retry boundaries, dry-run behavior, and provider-specific implementation modules.

Why MCP Beats Direct API Orchestration

Direct API integrations couple agent prompts to implementation details: OAuth quirks, endpoint payloads, retry rules, rate limits, and provider-specific naming. MCP moves those details into server-side tools. Claude asks for outcomes like fetch_high_intent_leads, send_followup_email, get_top_converting_pages, or create_iterable_journey; the MCP server owns validation, auth, logging, retries, provider transport, and safe fallbacks.

That separation makes the system easier to extend. A new lifecycle provider, ads channel, CRM field, or analytics source becomes a new tool module or transport implementation rather than a rewrite of every agent workflow.

MCP Architecture

autonomous-sales-system/
├── mcp_server/
│   ├── server.py                  # FastMCP entrypoint and tool registration
│   ├── config.py                  # Environment-driven provider configuration
│   ├── core/                      # Sales, CRM, Gmail, Sheets, WhatsApp, leads
│   ├── growth/                    # Ads, GA4, GSC, GTM, Contentful, Iterable
│   ├── infra/                     # Trigger.dev, webhooks, scheduling, retries
│   └── utils/                     # Auth, logging, validation, HTTP helpers
├── agents/                        # Agent policies and orchestration prompts
├── workflows/                     # Durable workflow and Trigger.dev mappings
└── tests/test_mcp_server.py        # MCP smoke test

Agent Layer

The repo now has an explicit agent registry in agents/registry.py with 30 business agents and 2 engines:

Outbound Engine: lead sourcing, qualification, personalization, outreach execution, follow-up strategy, meeting booking, CRM updates, and ICP refinement.
Inbound Engine: reply classification, inbound intent, revenue friction, churn prevention, expansion revenue, lifecycle automation, and CRM updates.

Claude can inspect and route agents through MCP tools: list_business_agents, get_business_agent, route_agent_task, and list_agent_engines.

Each business agent also has a provider API contract in agents/api_catalog.py. The MCP server exposes 30 agent-specific API tools, such as search_apollo_prospects, sync_hubspot_contact, create_calendly_invite, pull_stripe_revenue_events, pull_hotjar_friction_signals, and send_slack_founder_report. These tools run in dry-run mode by default and become live integrations when MCP_DRY_RUN=false and the required provider credentials are configured.

Available MCP Tool Surface

Core sales and CRM tools:

send_followup_email
search_gmail_replies
fetch_leads
fetch_high_intent_leads
create_or_import_lead
enrich_lead
update_google_sheet
send_whatsapp_message
update_pipeline_stage
get_campaign_performance

Growth, analytics, and lifecycle tools:

launch_google_ads_campaign
get_google_ads_performance
create_retargeting_campaign
pull_ga4_conversion_report
get_top_converting_pages
get_gsc_keyword_opportunities
get_seo_opportunities
validate_gtm_tracking_setup
analyze_gtm_tracking
update_contentful_page
optimize_contentful_landing_page
create_iterable_email_journey
create_iterable_journey

Infrastructure tools:

trigger_background_workflow
trigger_background_sales_workflow
get_background_job_status
schedule_workflow
list_scheduled_workflows
list_recent_webhook_events

How Claude Connects

Run the MCP server locally:

python -m mcp_server.server

Then configure Claude Desktop or your Claude MCP client to launch that command from the repository root. The server uses Anthropic's FastMCP interface from the Python MCP SDK and registers tools from each module during startup.

Local development defaults to MCP_DRY_RUN=true. That means Claude can exercise production-shaped tool calls without sending real emails, launching ads, editing Contentful, or messaging WhatsApp contacts. For production, set MCP_DRY_RUN=false and provide the provider credentials listed in .env.example.

Trigger.dev Role

Trigger.dev should own durable background work: lead enrichment, reply triage, multi-step follow-up sequences, lifecycle journey sync, campaign optimization, and periodic analytics jobs. Claude can call trigger_background_sales_workflow for long-running work instead of waiting inside a chat turn. The MCP server passes an idempotency key so retries do not accidentally duplicate outreach or campaign actions.

Recommended workflow mapping:

lead_enrichment: enrich, dedupe, score, and route new leads.
follow_up_sequence: schedule compliant email and WhatsApp steps with suppression rules.
reply_triage: classify intent, update CRM stage, and trigger next-best action.
growth_optimization: combine GA4, GSC, GTM, Ads, and Contentful signals.
lifecycle_journey_sync: create Iterable journeys for nurture and activation.

GTM Means Two Things

In this repo, GTM can mean:

Google Tag Manager: the tracking container used for tags, triggers, events, and conversion measurement. MCP tools like validate_gtm_tracking_setup and analyze_gtm_tracking refer to this meaning.
Go-To-Market: the broader sales and growth motion across leads, channels, campaigns, lifecycle, and revenue operations.

The system supports both: Google Tag Manager for measurement integrity, and Go-To-Market automation for revenue execution.

Production Deployment Strategy

Deploy the existing FastAPI backend and the MCP server as separate processes. Keep the backend responsible for REST endpoints, operator UI support, and CRM persistence. Keep the MCP server responsible for Claude tool access and provider integrations. Run Trigger.dev workers separately for durable jobs.

Production checklist:

Set MCP_DRY_RUN=false.
Store secrets in Railway, Render, Fly.io, AWS Secrets Manager, Doppler, or a similar secret manager.
Use OAuth/service-account credentials for Google Workspace, GA4, GSC, GTM, and Google Ads.
Add provider-specific rate limit handling inside the relevant tool module.
Route long-running or retry-sensitive operations through Trigger.dev.
Keep webhook endpoints signature-validated with WEBHOOK_SIGNING_SECRET.
Add structured logs and error reporting around every provider call.

Extending Tools

Add a new provider capability by creating or editing one module under mcp_server/core, mcp_server/growth, or mcp_server/infra, then registering it in mcp_server/server.py. Keep each tool business-oriented. Prefer create_retargeting_campaign over post_to_meta_endpoint, and prefer get_seo_opportunities over query_search_console_rows.

The MCP boundary should stay recruiter-grade and production-readable: Claude sees business actions; engineers see isolated provider transports, validation, retries, logging, and dry-run safety.

🎯 What It Does

Autonomous system that handles the entire sales funnel:

Lead Generation - Scrapes 5+ different data sources (YC, Hunter, LinkedIn, Apollo, Clearbit)
Outbound Messaging - AI-generated personalized emails/WhatsApp (with your approval)
Inbound Classification - Detects replies, classifies sentiment, suggests responses
Meeting Booking - Auto-schedules with Calendly
Lead Qualification - Autonomous qualification scoring
Follow-ups - Scheduled reminders with A/B testing
Metrics & Learning - Self-improving with weekly reports

💰 ROI

Metric	Value
Monthly Cost	$50-150
Replaces	$50K-60K/year (1 FTE sales rep)
Setup Time	2-3 weeks
Production Ready	4-5 weeks
ROI	100x+

🚀 Quick Start (5 minutes)

1. Clone & Setup

git clone <repo>
cd autonomous-sales-system
python -m venv venv
source venv/bin/activate  # or venv\Scripts\activate on Windows
pip install -r requirements.txt
cp .env.example .env

2. Get API Keys

CLAUDE_API_KEY=sk-...          # anthropic.com
SUPABASE_URL=https://...       # supabase.com
SUPABASE_KEY=eyJ...            # supabase
HUNTER_API_KEY=...             # hunter.io (optional)
APOLLO_API_KEY=...             # apollo.io
LINKEDIN_EMAIL=...             # for scraping
LINKEDIN_PASSWORD=...          # for scraping
GMAIL_CREDENTIALS=...          # Google Cloud Console
CALENDLY_API_KEY=...           # calendly.com
OPENAI_API_KEY=...             # for fallback (optional)

3. Run Lead Scraper

# Test with 20 leads
python scrapers/multi_source_scraper.py --limit 20 --sources yc,hunter

# Full run (all sources)
python scrapers/multi_source_scraper.py --workers 10

# By ICP (Ideal Customer Profile)
python scrapers/multi_source_scraper.py --icp "B2B SaaS" --limit 100

4. Start Backend

python -m uvicorn backend.api:app --reload --port 8000

If SUPABASE_URL and SUPABASE_KEY are set, the API uses Supabase. Without them, it runs with an in-memory store for local development and tests.

5. Start Frontend

cd frontend
npm install
npm start

Visit http://localhost:3000

Note: the backend API remains the strongest part of the repo, and the frontend now provides a usable operator console rather than a full CRM-grade product surface.

The current frontend now includes an operator console for:

reviewing pipeline metrics
creating leads manually
generating and approving drafts
queuing follow-ups
viewing job and activity history

Outbound email delivery now supports two modes:

EMAIL_DELIVERY_MODE=dry_run for safe local testing
EMAIL_DELIVERY_MODE=smtp for real sends through an SMTP provider such as Gmail or SendGrid

📁 Project Structure

autonomous-sales-system/
├── scrapers/                      # Lead generation engines
│   ├── yc_scraper.py            # YC directory
│   ├── hunter_scraper.py         # Hunter.io API
│   ├── apollo_scraper.py         # Apollo.io API
│   ├── linkedin_scraper.py       # LinkedIn (Selenium)
│   ├── clearbit_scraper.py       # Clearbit enrichment
│   └── multi_source_scraper.py   # Orchestrator (use this)
│
├── backend/
│   ├── api.py                    # FastAPI app (main server)
│   ├── models.py                 # Pydantic models (Lead, Draft, etc)
│   ├── database.py               # Supabase/SQLAlchemy
│   ├── agents/
│   │   ├── draft_agent.py        # Email/WhatsApp generation
│   │   ├── classifier_agent.py   # Reply sentiment analysis
│   │   ├── qualification_agent.py # Lead scoring
│   │   ├── booking_agent.py      # Calendar integration
│   │   └── gtm_agent.py          # Metrics & reporting
│   ├── integrations/
│   │   ├── gmail.py              # Gmail API
│   │   ├── calendly.py           # Calendly API
│   │   ├── whatsapp.py           # WhatsApp Business API
│   │   └── claude.py             # Claude API wrapper
│   └── jobs/
│       ├── scheduler.py          # APScheduler setup
│       ├── tasks.py              # Background jobs
│       └── queue.py              # Celery/Bull integration
│
├── frontend/
│   ├── src/
│   │   ├── App.jsx               # Main app
│   │   ├── pages/
│   │   │   ├── Dashboard.jsx     # Pipeline overview
│   │   │   ├── Outbound.jsx      # Draft review & approve
│   │   │   ├── Inbound.jsx       # Reply classifier
│   │   │   ├── Leads.jsx         # Lead management
│   │   │   └── Metrics.jsx       # Analytics & funnel
│   │   └── components/
│   │       ├── LeadCard.jsx
│   │       ├── DraftPanel.jsx
│   │       ├── ReplyClassifier.jsx
│   │       └── MetricsDashboard.jsx
│   └── package.json
│
├── config/
│   ├── icp_profiles.json         # Ideal Customer Profiles
│   └── scraper_config.yml        # Scraper settings
│
├── tests/
│   ├── test_scrapers.py
│   ├── test_api.py
│   └── test_agents.py
│
├── migrations/
│   ├── 001_initial_schema.sql    # Database schema
│   └── 002_add_indexes.sql
│
├── docs/
│   ├── ARCHITECTURE.md           # System design
│   ├── SETUP.md                  # Installation guide
│   ├── API.md                    # API reference
│   ├── BOTTLENECKS.md            # Performance issues & fixes
│   └── DEPLOYMENT.md             # Prod deployment
│
├── .env.example                  # Environment template
├── docker-compose.yml            # Local dev setup
├── Dockerfile                    # Production image
├── README.md                     # This file
└── requirements.txt              # Python dependencies

🔧 API Keys Setup

Anthropic (Claude)

Go to https://console.anthropic.com
Create API key
Add to .env: CLAUDE_API_KEY=sk-...

Supabase (Database)

Go to https://supabase.com
Create new project
Get credentials from Settings → API
Run migrations: psql -h $HOST -U $USER -d $DB < migrations/001_initial_schema.sql

Hunter.io (Email Enrichment)

Go to https://hunter.io
Sign up (free tier: 25 searches/month)
Get API key from dashboard

Apollo.io (B2B Leads)

Go to https://apollo.io
Sign up (free tier: 100 searches/month)
Get API key

LinkedIn (Scraping)

Create LinkedIn account
Add credentials to .env
Note: Use with caution (respects robots.txt)

Google (Gmail API)

Go to https://console.cloud.google.com
Create project
Enable Gmail API
Create OAuth 2.0 credentials
Download JSON, rename to google_credentials.json

Calendly

Go to https://calendly.com
Get API key from integrations
Add to .env

🎯 Ideal Customer Profiles (ICPs)

System supports multiple ICP configurations. Edit config/icp_profiles.json:

{
  "b2b_saas": {
    "keywords": ["SaaS", "B2B", "software"],
    "min_employees": 10,
    "max_employees": 5000,
    "industries": ["Software", "Technology", "B2B"],
    "job_titles": ["CEO", "Founder", "VP Sales", "Head of Growth"],
    "company_age_min": 1,
    "funding_min": 0,
    "locations": ["US", "EU"]
  },
  "fintech": {
    "keywords": ["fintech", "payments", "crypto", "finance"],
    "industries": ["Financial Services", "FinTech"],
    "job_titles": ["CEO", "CTO", "VP Product"],
    "min_employees": 5,
    "funding_min": 1000000
  },
  "ai_startups": {
    "keywords": ["AI", "ML", "machine learning", "LLM"],
    "industries": ["AI/ML", "Artificial Intelligence"],
    "job_titles": ["CEO", "Founder", "CTO"],
    "max_employees": 500,
    "company_age_min": 0,
    "company_age_max": 5,
    "funding_min": 0
  }
}

Run Scraper by ICP

# Scrape B2B SaaS companies
python scrapers/multi_source_scraper.py --icp b2b_saas --limit 100

# Scrape FinTech with specific sources
python scrapers/multi_source_scraper.py --icp fintech --sources hunter,apollo

# Scrape AI startups with 20 parallel workers
python scrapers/multi_source_scraper.py --icp ai_startups --workers 20 --limit 500

🚀 Deployment

Local Development

docker-compose up
# Starts: PostgreSQL, Redis, FastAPI backend, React frontend

Production (Railway)

Database

# Create Supabase project, run migrations
psql $SUPABASE_CONNECTION_STRING < migrations/001_initial_schema.sql

Backend

# Deploy to Railway
railway link
railway up

Frontend

# Deploy to Vercel
cd frontend
vercel --prod

Environment Variables (Production)

# .env.production
CLAUDE_API_KEY=sk-...
SUPABASE_URL=https://xxx.supabase.co
SUPABASE_KEY=eyJ...
HUNTER_API_KEY=...
APOLLO_API_KEY=...
DATABASE_URL=postgresql://...
REDIS_URL=redis://...
GMAIL_CREDENTIALS_JSON=...
CALENDLY_API_KEY=...
OPENAI_API_KEY=...

📊 Features

1. Multi-Source Lead Generation

✅ YC Directory (free, Algolia API)
✅ Hunter.io (25-100 searches/month)
✅ Apollo.io (100-500 searches/month)
✅ LinkedIn (Selenium scraper, caution advised)
✅ Clearbit (email enrichment)
✅ Parallel processing (5-20 workers)
✅ Smart caching (50-70% speedup)

2. Outbound Engine

✅ AI-generated drafts (Claude)
✅ Template caching (80-90% cache hit)
✅ Confidence scoring (auto-approve >8.0)
✅ Batch approval UI
✅ Email + WhatsApp support
✅ A/B testing variants

3. Inbound Pipeline

✅ Gmail watch + polling fallback
✅ Reply classification (sentiment)
✅ Objection extraction
✅ Auto-response suggestions
✅ Real-time dashboard

4. Automation

✅ Calendly meeting booking
✅ Auto follow-ups (3, 7, 14 days)
✅ Pre-call briefings
✅ Scheduled tasks (APScheduler/Celery)

5. Intelligence

✅ Lead scoring algorithm
✅ Funnel metrics (stages, conversion %)
✅ A/B test results tracking
✅ Weekly AI-generated reports
✅ Self-improving message templates

6. Observability

✅ Error tracking (Sentry)
✅ Job monitoring (Celery Flower)
✅ Database query logging
✅ API request logs
✅ Performance metrics

📈 Performance Targets

Metric	Before	After	Method
Lead generation (10K)	4-5 hours	30-45 mins	Parallel async + caching
Draft generation	3-5s	<100ms	Template caching
Reply detection	30-180s	5-10s	Watch + polling
Approval workflow	100 mins	5-10 mins	Batch + auto-approval
Database queries	10-30s	<100ms	Indexed Supabase
System uptime	85%	99.9%	Job queue + monitoring

🔐 Security

✅ No hardcoded secrets (use .env)
✅ API key rotation (monthly)
✅ Rate limiting (Supabase auth)
✅ CORS configured
✅ Input validation (Pydantic)
✅ SQL injection protection (SQLAlchemy ORM)
✅ OAuth 2.0 for Gmail

📝 Documentation

Document	Purpose
SETUP.md	Installation & configuration
ARCHITECTURE.md	System design & agents
API.md	REST API endpoints
BOTTLENECKS.md	Performance issues & solutions
DEPLOYMENT.md	Production deployment

🐛 Troubleshooting

Scraper Rate Limited

# Reduce workers
python scrapers/multi_source_scraper.py --workers 3

# Use proxy (paid)
export SCRAPER_API_KEY=...

Gmail Watch Not Detecting Replies

# Logs show: Check backend logs
docker logs autonomous-sales-system-api

# Solutions:
# 1. Verify Gmail credentials
# 2. Check polling fallback is enabled
# 3. Increase polling frequency (default 30s)

Claude API Over Budget

# Add caching
# Reduce tokens in prompts
# Enable batch generation (off-peak)
# See BOTTLENECKS.md for detailed solutions

Database Connection Errors

# Verify Supabase URL & key
# Check migrations ran: psql $URL < migrations/001_initial_schema.sql
# Verify tables exist: SELECT * FROM leads;

📚 Learning Resources

🤝 Contributing

We welcome contributions! Areas needing help:

Additional lead sources (Crunchbase, PitchBook)
Mobile app (React Native)
Email templates (more variants)
Integrations (Pipedrive, Hubspot)
Tests (unit, integration, E2E)

📄 License

MIT License - Use freely, modify, sell. See LICENSE file.

💡 Tips & Tricks

Maximize Lead Quality

Define strict ICP in icp_profiles.json
Use multiple sources (YC + Hunter + Apollo)
Enable email verification (Hunter)
Filter by funding/employee count

Improve Reply Rates

Start with warm leads (replies > 5%)
A/B test subject lines
Personalize with company-specific info
Follow up after 3-5 days

Reduce Costs

Cache aggressively (80%+ hit rate)
Batch lead imports (1x/day vs continuous)
Use free tier APIs first (Hunter 25/mo free)
Pre-generate drafts in background

Scale Safely

Test with 100 leads first
Monitor reply rate & costs daily
Start with 1 worker, scale to 10+
Use job queue for long-running tasks

🚦 Status

📞 Support

Issues: GitHub Issues
Discussions: GitHub Discussions
Email: hello@autonomous-sales.com

Ready to automate sales? Start with: python scrapers/multi_source_scraper.py --limit 20

Name		Name	Last commit message	Last commit date
.github/workflows		.github/workflows
agents		agents
backend		backend
config		config
docs		docs
frontend		frontend
mcp_server		mcp_server
migrations		migrations
scrapers		scrapers
tests		tests
workflows		workflows
.env.example		.env.example
.gitignore		.gitignore
Dockerfile		Dockerfile
LICENSE		LICENSE
PROJECT_SUMMARY.md		PROJECT_SUMMARY.md
QUICKSTART.md		QUICKSTART.md
README.md		README.md
docker-compose.yml		docker-compose.yml
requirements.txt		requirements.txt

Latest commit

History

Folders and files

Repository files navigation