Clay Workbench
guerrilla2799/clay-workbench
Automatize workflows Clay.com completos dentro do Claude Code
Ver no GitHubMCP server completo para construir, debugar e executar workbooks do Clay.com diretamente via Claude Code. Cobre toda a jornada de GTM B2B: limpeza de listas, enriquecimento multi-provider, scoring de ICP e intent, outbound personalizado, roteamento inbound em tempo real e auditoria de créditos. 16 sub-skills organizadas por estágio do workflow com execução MCP-first e fallback para UI manual.
Instalação
npx convez-ai guerrilla2799/clay-workbench1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
Casos de uso
- Construir listas ABM com dados enriquecidos e scoring automático
- Executar waterfalls de enriquecimento multi-provider com controle de créditos
- Rotear leads inbound em tempo real com scoring de intent composite
- Auditar portfólio de workbooks e otimizar gasto de créditos
- Gerar cópia outbound personalizada com validação de envio
Veredito

Ferramenta de nível enterprise para times de GTM B2B que já usam Clay e querem automação profunda via Claude. Excelente integração MCP-first, mas demanda expertise em workflows Clay e governança de créditos. Impacto real em eficiência de prospecção e qualidade de dados se usado com disciplina.
README
A tradução para o português deste README ainda está na fila.
clay-workbench
The complete Clay.com workbook plugin for Claude Code.
Build, debug, and run Clay workbooks end-to-end via Claude Code. Covers the full Clay workflow — pre-Clay list hygiene, ongoing CRM data hygiene, ABM list building, deep account + prospect research, multi-provider enrichment waterfalls, ICP fit scoring, composite buying-signal intent scoring, outbound copy with Sending Gate + sequencer push, real-time inbound routing, recurring trigger monitoring, Claygent prompt iteration, credit forecasting + provider comparison, portfolio-wide cost audit, single-workbook troubleshooting, and reusable workbook templates.
Hybrid router + 16 sub-skills. MCP-first execution with manual UI walkthrough fallback. Mandatory credit pre-flight. ICP-gates-before-credits enforcement. Client-voice detection.
What this gives you
16 sub-skills organized by Clay workflow stage:
🧹 Hygiene & data quality
| Slash Command | What it builds |
|---|---|
/clay-list-clean | Pre-Clay list hygiene — domain/email/title/geo normalization + dedup + suppression. Always run first on raw lists; cuts burn 30–60%. |
/clay-data-hygiene | Ongoing CRM-side hygiene — 3 linked tables (crm_record_health rolling freshness + crm_duplicates weekly dedup + refresh_queue prioritized re-enrichment) |
🏗️ Build
| Slash Command | What it builds |
|---|---|
/clay-abm-list | Account-keyed ABM workbook with firmographic + trigger gating |
/clay-account-research | Deep multi-source account briefs — priorities + news + hiring + leadership + competitive + entry-point hypothesis per company |
/clay-prospect-research | Deep person-level dossiers — POVs + activity + role scope + warm path + personalized entry line per contact |
/clay-enrich-waterfall | Multi-provider email + phone + LinkedIn waterfall, graceful exit |
🎯 Score
| Slash Command | What it builds |
|---|---|
/clay-icp-score | 10-point composite ICP fit score (firmographic + persona) with routing by band |
/clay-buying-signals | Composite multi-source intent score — fuses 8+ signal sources (intent data, web visits, content engagement, hiring, funding, tech adoption, leadership moves) into one time-decayed "now is the moment" score with mandatory backfill discrimination test |
🚀 Activate
| Slash Command | What it builds |
|---|---|
/clay-outbound | Personalized copy + Sending Gate + sequencer push, client-voice aware |
/clay-inbound-routing | Real-time webhook enrichment + scoring + routing + Slack alerts |
/clay-signal-monitor | Recurring single-source trigger monitoring (job changes / hiring / funding / leadership / tech adoption / intent) → severity classify → Slack/CRM routing |
🔍 Quality & cost
| Slash Command | What it builds |
|---|---|
/clay-claygent-iterator | Structured 5 → 25 → 100 row Claygent prompt iteration loop with scorecard (PASS %, FABRICATED %, credits/row). Mandatory before scaling any Claygent column. |
/clay-credits | Proactive credit forecasting + multi-mix provider cost comparison + workbook ROI ranking with reallocation rec |
/clay-cost-audit | Portfolio-wide 12-check audit across every workbook (Global Rules violations + waste + ROI laggards) with executive / engineering / Slack output modes |
/clay-troubleshoot | Diagnose broken / expensive / under-performing workbooks (single-workbook reactive) |
💾 Persistence
| Slash Command | What it builds |
|---|---|
/clay-template-library | Save / load / list / diff / version / share workbook templates with anonymization rules. Catalog at templates/library/INDEX.md. |
🧭 Router
| Slash Command | What it builds |
|---|---|
/clay | Master router — auto-detects intent, routes to the right sub-skill |
Why it's different
Most Clay skills are either "consultancy markdown" (here are the steps) or "raw scripts" (here is the JSON). This one is both — and more:
- MCP-first execution. Calls
mcp__claude_ai_Clay__*tools directly to actually run workbooks. Falls back to step-by-step UI walkthrough when MCP can't do it yet. - Mandatory credit pre-flight. Estimates
rows × credits/rowagainst your current balance before ANY paid run > 100 rows. Blocks at 50% balance threshold by default. - Two-pass ICP gate as a global rule. Cheap signals first (industry, size, geo); paid enrichments gated on pass 1. Typically cuts burn 40–60%.
- Strict Sending Gate. Sequencer push requires
sending_gate_eligible = TRUEformula column. Per-step eligibility (step 1, step 2, LinkedIn) supported out of the box. - Client voice detection. When the prompt mentions a known client (Obin, Vantage, etc.), routes to the matching messaging skill for ICP + voice. Generic otherwise.
- 3 golden reference workbooks. Production-ready templates — copy/paste into Clay.
- Action registry with gotchas. The classic
actionKey: "ai"silent-drop bug? Documented and avoided.
The 8 Global Rules (Non-Negotiable)
Every sub-skill enforces these:
- Gates Before Credits — ICP gate column required before any paid column
- Sending Gate Before Export — formula column required before any sequencer push
- Data Unification — score on full rows, not fragments
- ICP Qualification First — cheapest-signal disqualification before enrichment
- Free Before Paid — try free sources first in every waterfall
- Never Assume — Always Ask — run the 8-question intake; offer options when user defers
- Clay-Managed vs Own API Key — state which side each column hits
- Multiple Sending Steps — outbound assumes 2–4 step cadence, not one-shot
Violation requires explicit gate-override: <reason> declaration.
Installation
Option A — Clone + register all 16 slash commands
git clone https://github.com/guerrilla2799/clay-workbench ~/.claude/claudecodeskills/clay-workbench
Then register each slash command:
SKILLS=(
clay-list-clean
clay-data-hygiene
clay-abm-list
clay-account-research
clay-prospect-research
clay-enrich-waterfall
clay-icp-score
clay-buying-signals
clay-outbound
clay-inbound-routing
clay-signal-monitor
clay-claygent-iterator
clay-credits
clay-cost-audit
clay-troubleshoot
clay-template-library
)
# Master router
cat > ~/.claude/commands/clay.md <<'EOF'
---
allowed-tools: Read, Write, Edit, WebSearch, WebFetch, AskUserQuestion, Glob, Grep, Bash
description: clay-workbench — master router
---
$ARGUMENTS
See ~/.claude/claudecodeskills/clay-workbench/SKILL.md for full instructions.
EOF
# Sub-skills
for skill in "${SKILLS[@]}"; do
cat > ~/.claude/commands/${skill}.md <<EOF
---
allowed-tools: Read, Write, Edit, WebSearch, WebFetch, AskUserQuestion, Glob, Grep, Bash
description: clay-workbench — ${skill}
---
\$ARGUMENTS
See ~/.claude/claudecodeskills/clay-workbench/skills/${skill}/SKILL.md for full instructions.
EOF
done
Option B — Install as a Claude Code plugin
The .claude-plugin/plugin.json manifest is included for plugin-manager-based installs. Drop the folder into your Claude Code plugins directory.
Required: Clay MCP server
This plugin assumes the official Clay MCP server is connected (mcp__claude_ai_Clay__* tools). If not:
- Connect at https://claude.ai/settings/connectors (or your equivalent Claude Code MCP config)
- Authenticate with your Clay workspace
- Verify with
mcp__claude_ai_Clay__list_subroutines— should return your subroutines
Without the MCP server, the plugin falls back to manual UI walkthroughs (which work but lose live execution).
Quick start
# Pick your motion. Auto-router will pick the right sub-skill if you're unsure.
/clay
# Or jump to a specific sub-skill (sampling across stages):
# Hygiene
/clay-list-clean dedupe + suppress my Apollo export before any paid enrichment
/clay-data-hygiene stand up ongoing dedup + decay detection on my HubSpot CRM —
daily freshness scan, weekly dedup, rolling refresh queue
# Build
/clay-abm-list build a Tier 1 ABM list of B2B SaaS companies with 500-2500 employees
that closed a Series A or B in the last 90 days
/clay-account-research generate executive briefs for my Tier 1 list —
priorities + news + hiring + entry-point hypothesis
/clay-prospect-research deep dossier on these 25 named champions, with personalized
first-line drafts ready to hand to /clay-outbound
/clay-enrich-waterfall add an email + phone waterfall to my contacts table —
budget is 10¢/row, EU-heavy list, need 80% match rate
# Score
/clay-icp-score build a 10-point fit score for my MQL routing —
back-test against last 12 months closed-won
/clay-buying-signals fuse 6sense + Bombora + web reveal + hiring + funding into
one composite intent score — discriminate vs closed-won/lost
# Activate
/clay-outbound generate 4-step cold outbound copy from my T1 contacts table,
push to SalesLoft, voice is Obin AI
/clay-inbound-routing wire up real-time enrichment + routing on my HubSpot demo form
/clay-signal-monitor stand up a daily job-change watcher across my Champion alumni
list with severity-tiered Slack routing
# Quality & cost
/clay-claygent-iterator tune this Claygent prompt before I scale to 500 rows —
run the 5/25/100 loop, give me the scorecard
/clay-credits should I add ZoomInfo own-key, or upgrade Clay tier? Show me the
math on 50k contacts/quarter at current burn
/clay-cost-audit monthly portfolio audit — run all 12 checks across every
workbook, rank fixes by savings
/clay-troubleshoot my email waterfall is burning $400/day with 30% match rate —
diagnose and fix
# Persistence
/clay-template-library save the Obin Tier 1 ABM workbook as an anonymized template
for reuse across other clients
Plugin structure
clay-workbench/
├── SKILL.md # Master router
├── .claude-plugin/
│ └── plugin.json # Formal plugin manifest (v2.0.0)
├── skills/ # 16 sub-skills
│ ├── clay-list-clean/SKILL.md # Hygiene
│ ├── clay-data-hygiene/SKILL.md
│ ├── clay-abm-list/SKILL.md # Build
│ ├── clay-account-research/SKILL.md
│ ├── clay-prospect-research/SKILL.md
│ ├── clay-enrich-waterfall/SKILL.md
│ ├── clay-icp-score/SKILL.md # Score
│ ├── clay-buying-signals/SKILL.md
│ ├── clay-outbound/SKILL.md # Activate
│ ├── clay-inbound-routing/SKILL.md
│ ├── clay-signal-monitor/SKILL.md
│ ├── clay-claygent-iterator/SKILL.md # Quality & cost
│ ├── clay-credits/SKILL.md
│ ├── clay-cost-audit/SKILL.md
│ ├── clay-troubleshoot/SKILL.md
│ └── clay-template-library/SKILL.md # Persistence
├── resources/
│ ├── global-rules.md
│ ├── intake-questions.md
│ ├── mcp-tool-map.md
│ ├── action-registry.md
│ ├── credit-cost-table.md
│ ├── client-context-detection.md
│ ├── clayscript-library.md
│ ├── ai-prompt-templates.md
│ ├── enrichment-benchmarks.md
│ ├── claygent-guide.md
│ └── workflow-patterns.md
├── providers/ # Per-provider notes
│ ├── enrichment/ # Apollo, Findymail, Datagma, Dropcontact, ZeroBounce, Hunter, PDL, Clearbit, ContactOut
│ ├── databases/ # Sales Nav, Ocean.io
│ ├── signals/ # Predictleads, job changes, hiring/funding
│ ├── crm/ # HubSpot, Salesforce
│ └── sequencers/ # SalesLoft, Outreach, Apollo, 11x, Smartlead/Instantly
└── templates/ # Reference workbooks + reusable library
├── 01-abm-tier1-with-triggers.md
├── 02-inbound-demo-enricher-router.md
├── 03-cold-outbound-multi-provider-waterfall.md
└── library/ # Versioned templates from /clay-template-library
└── INDEX.md # Registry of saved templates
Auto-generated cross-template composition graph + per-template column DAGs live in docs/composition/ (regenerate via python3 scripts/compose-graph.py).
Design provenance
Composed from the best patterns across the Claude Code Clay ecosystem:
| Pattern | Source |
|---|---|
| Master + sub-skill split with routing table | forma-norden/clay-claude-code-skill-pack |
| 8 Global Rules (Gates Before Credits, etc.) | mariosworkflows/clay-engineer |
| 8-question mandatory intake | LirKonu/clay-campaign |
| Action registry + verification pattern | TenSpy-ai/claycast |
| Pre-built workflow patterns | sachacoldiq/ColdIQ-s-GTM-Skills |
| Standalone + cross-link footer | ColdIQ |
The four load-bearing patterns: forma's routing table, clay-engineer's Global Rules, clay-campaign's 8-Q intake, claycast's action registry.
What it WON'T do
- Build a workbook without running the 8-question intake first
- Add paid columns before an ICP gate exists
- Push to a sequencer without a Sending Gate column
- Run anything with
auto_run = trueduring construction - Generate outbound copy without confirming voice / client context
- Skip the credit pre-flight on runs > 100 rows
These are by design. Override per case with gate-override: <reason>.
Roadmap
v2.0.0 (this release):
- Subroutine save / load — shipped as
/clay-template-librarywith anonymization rules, versioning, INDEX.md catalog - Cost analytics dashboard — shipped as
/clay-credits(forecast + provider comparison + workbook ROI) and/clay-cost-audit(portfolio-wide 12-check audit) - Claygent prompt iteration loop — shipped as
/clay-claygent-iterator(5 → 25 → 100 row loop with scorecard) - Backfill pattern: detect existing workbooks missing Global Rules — shipped as
/clay-cost-audit - Deep research layer — shipped as
/clay-account-research+/clay-prospect-research - Monitoring layer — shipped as
/clay-signal-monitor(single source) +/clay-buying-signals(multi-source composite) - Data hygiene layer — shipped as
/clay-list-clean(pre-Clay) +/clay-data-hygiene(ongoing CRM)
Post-v2.0.0 polish + roadmap (shipped 2026-06-10):
-
resources/global-rules.mdcross-refs to all 10 new sub-skills under each rule -
templates/library/seeded with 10 bootstrap templates (abm-account-keyed-tier-1,account-research-tier-1-brief,email-waterfall-eu,email-waterfall-us-smb,inbound-router-demo-form,outbound-3-step-cadence-cold,outbound-3-step-cadence-warm,prospect-research-champion-brief,signal-monitor-hiring-posture,signal-monitor-job-change) + stage-grouped INDEX.md catalog - Subroutine export — EXPORT mode added to
/clay-template-libraryfor automated round-trip from a live Clay subroutine or table → portabletemplate.jsonwith workspace-portability transforms + anonymization applied - Community templates via PR —
CONTRIBUTING.md+.github/PULL_REQUEST_TEMPLATE.md+scripts/validate-template.py(stdlib only; schema + anonymization checks) - Cross-workbook composition visualizer —
scripts/compose-graph.pyreadstemplates/library/*/template.jsonand emits Mermaid graphs atdocs/composition/library-graph.md+docs/composition/per-template/<slug>.md; idempotent with--checkflag for CI drift detection
Next:
- Schema operations — create new tables via MCP when Clay exposes the API. Blocked on Clay exposing a write API for tables; out of our hands until then.
Contributing
PRs welcome — especially anonymized templates for templates/library/. See CONTRIBUTING.md for the template contribution workflow, the anonymization checklist, and the local validator script.
License
MIT — see LICENSE.
Author
Brandon Redlinger — Fractional VP Marketing / GTM engineer.
Built for the Stack & Scale audience: GTM operators who think in systems, not tactics.