B2B SDR Agent Template

We're live on Product Hunt! → PulseAgent B2B SDR Agent Skill — Open-source AI SDR for WhatsApp, Email & Telegram. An upvote means the world to us 🙏

Turn any B2B export business into an AI-powered sales machine in 5 minutes.

✅ New · 2026-06-07 — Template v3.8.0: GitHub Actions CI for npm test, post-deploy deploy/doctor.sh health checks, configurable remote OpenClaw paths for non-root deployments, and expanded validation for skill counts + workspace path generation. See full changelog →

🚀 New · 2026-06-12 — OpenClaw v2026.6.6: Security hardening across 10+ vectors (exec approvals fail-closed, deleted-agent ACP bypasses closed), Telegram account-scoped topic routing, iMessage always-on inbound restart, WhatsApp successor-controller fix, Claude Fable 5 adaptive thinking support, and faster UI startup via cached model metadata. See full changelog →

📦 Add-on · 2026-05-22 — WhatsApp Legacy Account Onboarding Spec v0.5 (Path D — Multi-Device fetch): Lowest-friction history acquisition — customer scans one QR, waits for Multi-Device sync, bootstrap.sh pulls .txt files straight from PulseAgent. No backup, no password, no USB. Backup-extraction (v0.4) still default for zero-downtime tenants. Three-layer architecture (MemOS + sales_playbook + conversation_history), A/B/C/D delivery decision tree, strict auto-onboard gate. Release notes →

An open-source, production-ready template for building AI Sales Development Representatives (SDRs) that handle the full sales pipeline — from lead capture to deal closing — across WhatsApp, Telegram, and email.

Built on OpenClaw, battle-tested with real B2B export companies.

🌐 English | 中文 | Español | Français | العربية | Português | 日本語 | Русский

---

Architecture: 7-Layer Context System

┌─────────────────────────────────────────────────┐
│              AI SDR Agent                       │
├─────────────────────────────────────────────────┤
│  IDENTITY.md   → Who am I? Company, role        │
│  SOUL.md       → Personality, values, rules     │
│  AGENTS.md     → Full sales workflow (10 stages)│
│  USER.md       → Owner profile, ICP, scoring    │
│  HEARTBEAT.md  → 14-item pipeline inspection    │
│  MEMORY.md     → 4-layer anti-amnesia protocol  │
│  TOOLS.md      → CRM, channels, integrations    │
├─────────────────────────────────────────────────┤
│  Skills        → Extensible capabilities        │
│  Product KB    → Your product catalog           │
│  Cron Jobs     → 14 automated pipeline checks   │
├─────────────────────────────────────────────────┤
│  OpenClaw Gateway (WhatsApp / Telegram / Email) │
└─────────────────────────────────────────────────┘

Each layer is a Markdown file you customize for your business. The AI reads all layers on every conversation, giving it deep context about your company, products, and sales strategy.

Quick Start

Option A: OpenClaw Users (1 Command)

If you already have OpenClaw running:

clawhub install b2b-sdr-agent

Browse more skills on ClawHub — the AI agent skill marketplace.

Done. The skill installs the full 7-layer context system, delivery-queue, and sdr-humanizer into your workspace. Then customize:

# Edit the key files for your business
vim ~/.openclaw/workspace/skills/b2b-sdr-agent/references/IDENTITY.md
vim ~/.openclaw/workspace/skills/b2b-sdr-agent/references/USER.md

# Or copy to your main workspace
cp ~/.openclaw/workspace/skills/b2b-sdr-agent/references/*.md ~/.openclaw/workspace/

Replace all {{placeholders}} with your actual company info, and your AI SDR is live.

Option A2: Hermes Agent Users (1 Command)

If you use Hermes Agent instead of OpenClaw:

hermes skills install github:iPythoning/b2b-sdr-hermes-skill

Same 10-stage SDR pipeline, adapted for Hermes conventions (native cron, skill config, built-in WhatsApp). Migrating from OpenClaw? Run hermes claw migrate first, then install the skill.

→ b2b-sdr-hermes-skill repo

Option B: Full Deployment (5 Minutes)

1. Clone & Configure

git clone https://github.com/iPythoning/b2b-sdr-agent-template.git
cd b2b-sdr-agent-template

# Edit the 7 workspace files for your business
vim workspace/IDENTITY.md   # Company info, role, pipeline
vim workspace/USER.md       # Your products, ICP, competitors
vim workspace/SOUL.md       # AI personality and rules

2. Set Up Deployment Config

cd deploy
cp config.sh.example config.sh
vim config.sh               # Fill in: server IP, API key, WhatsApp number

3. Validate Locally

cd ..
npm test

This checks the template structure, shell scripts, product KB JSON, proforma invoice generation, Chroma memory smoke flow, and generated OpenClaw config before you touch a server.

4. Deploy

cd deploy
./deploy.sh my-company

# Output:
# ✅ Deploy Complete: my-company
# Gateway:  ws://your-server:18789
# WhatsApp: Enabled
# ChromaDB Memory: Enabled (chromadb + local chroma-memory)
# Skills:   b2b_trade (41 skills)

5. Verify Remote Health

./doctor.sh

Doctor checks SSH, Node/npm, OpenClaw CLI, openclaw.json permissions, workspace context files, local skills, Gateway service state, Gateway /health, and Chroma memory.

That's it. Your AI SDR is live on WhatsApp and ready to sell.

What It Does

Full-Pipeline Sales Automation (10 Stages)

Stage	What the AI Does
1. Lead Capture	Auto-detect inbound messages (WhatsApp/Telegram/CTWA ads), create CRM records
2. BANT Qualification	Natural conversation to assess Budget, Authority, Need, Timeline
3. CRM Entry	Structured data capture — name, company, country, ICP score, product interest
4. Research & Enrichment	Jina AI web search + company website analysis, 3-layer enrichment pipeline
5. Quotation	Auto-generate PDF quotes, multi-language, send to owner for approval
6. Negotiation	Track counter-offers, recommend strategy, escalate when needed
7. Reporting	Daily 09:00 pipeline reports, 15:00 stalled alerts, weekly summaries
8. Nurture	Automated follow-ups, industry news, post-sale care, quarterly check-ins
9. Email Outreach	Personalized cold email sequences (Day 1/3/7/14), auto follow-up
10. Multi-Channel Orchestration	Cross-channel coordination (WhatsApp + Email + Telegram), auto-switching

Automated Scheduling (14 Pipeline Checks)

Schedule	Task
Every heartbeat	New lead intake, quote tracking, meeting prep, and data-quality checks
Every 30 min	Gmail inbox scan for client replies
Daily 09:00	Pipeline report to owner via WhatsApp
Daily 10:00	AI lead discovery (market rotation: Africa/ME/SEA/LatAm)
Daily 11:00	Email follow-up sequence check
Daily 15:00	Stalled lead detection
Weekly Wed	Nurture campaign
Weekly Fri	Competitor intelligence gathering
Weekly Mon	Weekly summary report
Daily 12:00	CRM snapshot to ChromaDB (L4 disaster recovery)
Daily 14:00	Memory health check (Supermemory + ChromaDB stats)
Every heartbeat	Microsoft Teams mention/DM scan for sales handoff
Every 30 min	WhatsApp 72h window expiry check → auto-switch to Telegram

Human-Like Conversations

• Responds in the customer's language (auto-detected)
• Natural message pacing with delivery-queue skill (3-90s delays, message splitting)
• Timezone-aware sending (7 market timezone mappings)
• Cultural adaptation (Middle East, Africa, SEA, LatAm, Europe)
• Never reveals AI identity — presents as your sales consultant

Dynamic ICP Scoring

• Initial score based on 5 weighted dimensions (purchase volume, product match, region, payment ability, authority)
• Auto-adjusts based on interaction: fast reply +1, asks for quote +2, mentions competitor +2, 7d no reply -1
• Hot leads (ICP>=7) auto-flagged, owner notified immediately

4-Layer Anti-Amnesia System

AI agents lose context in long conversations and across sessions. Our 4-layer anti-amnesia architecture ensures your AI SDR never forgets:

Message In ──→ L1 MemOS auto-recall (structured memory injection)
    │
    ├──→ L3 ChromaDB per-turn store (customer-isolated, auto-tagged)
    │
    ├──→ L2 Proactive Summary at 65% tokens (haiku compression, zero info loss)
    │
    └──→ L4 CRM Snapshot daily 12:00 (disaster recovery fallback)

Layer	Engine	What It Does
L1: MemOS	Structured memory	Auto-extracts BANT, commitments, objections every turn. Injects into System Prompt at conversation start.
L2: Proactive Summary	Token monitoring	Compresses at 65% context usage via haiku-class model. All numbers, quotes, commitments preserved verbatim.
L3: ChromaDB	Per-turn vector store	Every conversation turn stored with customer_id isolation. Auto-tags quotes, commitments, objections. Semantic retrieval across sessions.
L4: CRM Snapshot	Daily backup	Stores full pipeline state daily to ChromaDB as disaster recovery. If any layer fails, L4 has the data.

Result: Your AI SDR remembers every customer, every quote, every commitment — even after 100+ turns, weeks of silence, or system restarts.

See ANTI-AMNESIA.md for the full implementation spec with code, prompts, and deployment guide.

The 7 Layers Explained

Layer	File	Purpose
Identity	IDENTITY.md	Company info, role definition, pipeline stages, lead tiering
Soul	SOUL.md	AI personality, communication style, hard rules, growth mindset
Agents	AGENTS.md	10-stage sales workflow, BANT qualification, multi-channel orchestration
User	USER.md	Owner profile, product lines, ICP scoring, competitors
Heartbeat	HEARTBEAT.md	14-item automated pipeline inspection + memory health
Memory	MEMORY.md	4-layer anti-amnesia protocol (MemOS + Summary + ChromaDB + CRM)
Tools	TOOLS.md	CRM commands, channel config, web research, email access

Skills

Pre-built capabilities that extend your AI SDR:

Skill	Description
delivery-queue	Schedule messages with human-like delays. Drip campaigns, timed follow-ups.
supermemory	Semantic memory engine. Store research notes, competitor intel, market insights.
chroma-memory	Per-turn conversation storage with customer isolation, auto-tagging, and CRM snapshots.
telegram-toolkit	Bot commands, inline keyboards, large file handling, and Telegram-first market strategies.
sdr-humanizer	Rules for natural conversation — pacing, cultural adaptation, anti-patterns.
lead-discovery	AI-driven lead discovery. Web search for potential buyers, ICP evaluation, CRM auto-entry.
quotation-generator	Auto-generate PDF proforma invoices with company letterhead, multi-language support.
graphify	Knowledge graph engine — map product relationships, customer intelligence, and market research into queryable graphs. Powered by graphify.

Skill Profiles

Choose a pre-configured skill set based on your needs:

Profile	Skills	Best For
b2b_trade	41 skills	B2B export companies (default)
lite	16 skills	Getting started, low-volume
social	14 skills	Social media-focused sales
full	40+ skills	Everything enabled

Industry Examples

Ready-to-use configurations for common B2B export verticals:

Industry	Directory	Highlights
Heavy Vehicles	examples/heavy-vehicles/	Trucks, machinery, fleet sales, African/ME markets
Consumer Electronics	examples/electronics/	OEM/ODM, Amazon sellers, sample-driven sales
Textiles & Garments	examples/textiles/	Sustainable fabrics, GOTS certified, EU/US markets

To use an example, copy it into your workspace:

cp examples/heavy-vehicles/IDENTITY.md workspace/IDENTITY.md
cp examples/heavy-vehicles/USER.md workspace/USER.md
# Then customize for your specific business

Product Knowledge Base

Structure your product catalog so the AI can generate accurate quotes:

product-kb/
├── catalog.json                    # Product catalog with specs, MOQ, lead times
├── products/
│   └── example-product/info.json   # Detailed product info
└── scripts/
    └── generate-pi.js              # Proforma invoice generator

Control Dashboard

After deployment, your AI SDR comes with a built-in web dashboard:

http://YOUR_SERVER_IP:18789/?token=YOUR_GATEWAY_TOKEN

The dashboard shows:

• Real-time bot status and WhatsApp connection
• Message history and conversation threads
• Cron job execution status
• Channel health monitoring

The token is auto-generated during deployment and printed in the output. Keep it private — anyone with the URL+token has full access.

Security note: Set GATEWAY_BIND="loopback" in config.sh to disable remote dashboard access. Default is "lan" (accessible from network).

WhatsApp Configuration

By default, the AI SDR accepts messages from all WhatsApp contacts (dmPolicy: "open"). This is the recommended setting for sales agents — you want every potential customer to be able to reach you.

Setting	Value	Meaning
WHATSAPP_DM_POLICY	"open" (default)	Accept DMs from anyone
	"allowlist"	Only accept from ADMIN_PHONES
	"pairing"	Require pairing code first
WHATSAPP_GROUP_POLICY	"allowlist" (default)	Only respond in whitelisted groups

To change after deployment, edit ~/.openclaw/openclaw.json on the server:

{
  "channels": {
    "whatsapp": {
      "dmPolicy": "open",
      "allowFrom": ["*"]
    }
  }
}

Then restart: systemctl --user restart openclaw-gateway

💡 Keep your sender number alive: aggressive outreach gets new WhatsApp numbers banned fast. Free toolkit at PaibaoWork （中文） — ban self-check, account warm-up tracker, appeal letter generator, team anti-association check.

Deployment

Prerequisites

• A Linux server (Ubuntu 20.04+ recommended)
• Node.js 18+
• An AI model API key (OpenAI, Anthropic, Google, Kimi, etc.)
• WhatsApp Business account (optional but recommended)

Configuration

All configuration lives in deploy/config.sh. Key sections:

# Server
SERVER_HOST="your-server-ip"

# AI Model
PRIMARY_API_KEY="sk-..."

# Channels
WHATSAPP_ENABLED=true
TELEGRAM_BOT_TOKEN="..."

# CRM
SHEETS_SPREADSHEET_ID="your-google-sheets-id"

# Admin (who can manage the AI)
ADMIN_PHONES="+1234567890"

For non-root servers or custom OpenClaw locations, set the remote paths instead of editing deploy scripts:

SERVER_USER="openclaw"
REMOTE_HOME="/srv/openclaw"
REMOTE_OPENCLAW_HOME="/srv/openclaw/.openclaw"
REMOTE_CONFIG_HOME="/srv/openclaw/.config"
REMOTE_WORKSPACE_DIR="/srv/openclaw/.openclaw/workspace"

CI runs npm test on every push and pull request to main.

WhatsApp IP Isolation (Multi-Tenant)

When running multiple agents on the same server, each should have a unique exit IP so WhatsApp sees independent devices. This prevents cross-account flagging.

# After deploying a client, isolate their WhatsApp IP:
./deploy/ip-isolate.sh acme-corp

# Or with a specific SOCKS5 port:
./deploy/ip-isolate.sh acme-corp 40010

How it works:

                  ┌─ wireproxy :40001 → WARP Account A → CF IP-A
                  │    ↑
tenant-a ─────────┘    ALL_PROXY=socks5://host:40001

tenant-b ─────────┐    ALL_PROXY=socks5://host:40002
                  │    ↓
                  └─ wireproxy :40002 → WARP Account B → CF IP-B

Each tenant gets:

• A dedicated free Cloudflare WARP account
• An isolated wireproxy instance (~4MB RAM)
• A unique Cloudflare exit IP for all outbound traffic (including WhatsApp)

To auto-enable during deploy, set IP_ISOLATE=true in config.sh.

Managed Deployment

Don't want to self-host? PulseAgent offers fully managed B2B SDR agents with:

• One-click deployment
• Dashboard & analytics
• Multi-channel management
• Priority support

Get Started →

Community

Join the conversation, ask questions, and share your use cases:

• Discord — Real-time chat with the team
• Telegram — Community discussions
• X / Twitter — Announcements & updates
• Product Hunt — Upvote & reviews

Learn more:

• PulseAgent Blog — Tutorials, release notes, and case studies
• Open Source Hub — All repos, docs, and ecosystem links
• ClawHub — Browse and install AI agent skills
• Running on Kimi K2.6 — LLM setup guide (pitfalls, auth profiles, context budgeting)

Contributing

Contributions welcome! See our Contributing Guide for details. Areas where we'd love help:

• Industry templates: Add examples for your industry
• Skills: Build new capabilities
• Translations: Translate workspace templates to other languages
• Documentation: Improve guides and tutorials

License

MIT — use it for anything.

---

---

TalentDelta provenance record (added at packaging)

• Packaged for: TalentDelta catalogue (draft; gated by review)
• Upstream: https://github.com/iPythoning/b2b-sdr-agent-template
• Pinned commit: 6131928188d863fb413b70f3da8b6d3e13f44eea (2026-09-29)
• Licence: MIT - the upstream LICENSE file ships unchanged
• Upstream author credit: Clark Fan (iPythoning / PulseAgent)
• Runtime: OpenClaw agent runtime (this template does not target Hermes): Node 18+ for install.sh and the package.json tooling, Python for the scripts, a messaging channel (WhatsApp / Telegram / email) with its own credentials, and an LLM API key. deploy/generate-config.sh builds the runtime config from deploy/config.sh.example
• Changes made at packaging: this section was appended to README.md. Upstream paths deliberately NOT shipped in this archive: .sync (the maintainer's own blog and WeChat publishing state (draft posts, hourly sync logs, release markers) - regenerable, and unrelated to running the agent). Edited at packaging, so NOT byte-identical to the pinned commit: scripts/validate-template.sh, skills/chroma-memory/SKILL.md, skills/delivery-queue/SKILL.md, skills/sdr-humanizer/SKILL.md, skills/supermemory/SKILL.md, skills/telegram-toolkit/SKILL.md (five capabilities shipped without a YAML frontmatter block, so no loader could read them: a block was inserted with name set to the skill directory and description taken verbatim from that skill's own summary line, leaving the original heading and body untouched; in the deployment validation harness two values were credential-shaped (written into a temporary config and read back in a second phase) and are now REPLACE_WITH_YOUR_REAL_KEY, which keeps the test meaningful and leaves nothing secret-shaped in the archive. No other behaviour changed; bash -n passes on the edited script). Every other file is byte-identical to the pinned commit.
• Safety read of the packaged archive (packaging time, 2026-10-06): no live credentials. The only credential-shaped values were two test values in scripts/validate-template.sh, replaced at packaging with a neutral placeholder; the harness still parses (bash -n clean). See the packaging notes below.

Packaging notes (TalentDelta)

• The reviewer-side inspector's two remaining findings are disclosed rather than papered over. scripts/validate-template.sh:101 writes the template's own documented non-root deployment home /home/openclaw and asserts the generated config against it: it is the TEMPLATE's path, not the packager's, and the assertion fails if it is changed. README.md tells the operator the Doctor checks openclaw.json; that file is deployment output produced by deploy/generate-config.sh from deploy/config.sh.example (both shipped), and the upstream repository does not contain it either.
• Repairs verified before/after by sha256 prefix: chroma-memory 3979a27f->8a59a61c, delivery-queue b36ae369->efec20cc, sdr-humanizer 75a13a93->4711cdaf, supermemory 210c8ac0->8960a504, telegram-toolkit d709830b->a3d3bac0, scripts/validate-template.sh c58ea711->251ed99f.
• Language: the workflow document AGENTS.md is written in Chinese; workspace/ identity files, docs/, deploy/ and the multilingual README suite are English (README in 8 languages).
• The root SKILL.md declares name: b2b-sdr-agent while its directory is b2b-sdr-agent-template - an upstream cosmetic mismatch, left as-is and reported by the inspector as a warning.