active
CappaWork
A whole consultancy, run by one operator and Claude.
I needed a CRM, a pipeline, a client portal, and a task tracker. I also needed a way to talk to it when a good idea hit while walking my girls to school. Six tools that don’t talk to each other wouldn’t cut it, so I built one platform that does it all.
Claude runs the prospect research and the pipeline work. The public site is built to be found by AI agents, not just people. And I run the whole thing by voice, from my phone, through an MCP server.
At a glance
I run a premium AI consultancy alone. Lead generation, qualification, contracts, delivery, and the CRM all land on me — and I have to stay legible to the AI assistants buyers now use to find vendors, not just to people.
So I built one Next.js platform: a marketing site, client delivery portals, and an AI back office where Claude drives the business-development pipeline through tool use. The public surface is engineered for AI agents, and a companion MCP server lets me run the CRM by voice.
The hard part was putting an LLM into my system of record — with tools that mutate a live pipeline — without letting it go rogue. I drew a hard trust boundary: strict schemas, scoring computed in code, and human review before any record commits, with authorization enforced in app code rather than at the database.
It collapses five or six separate SaaS tools and a sales assistant into one system I run end to end — from first touch through signed contract to delivered project.
AI components
Claude Sonnet 4 for prospect/deal extraction, web-grounded enrichment, and command orchestration; Claude as the MCP client.
An admin command bar that runs a bounded tool-use loop over ~8 typed tools to act on the live BD pipeline.
None — deliberately uses Claude’s native web search for real-time prospect intelligence instead of maintaining a vector index.
None yet — correctness rests on strict JSON schemas and human review of every parsed record before it commits.
Text-to-record extraction pipelines (profiles, emails, notes → validated JSON → Postgres) with deterministic priority scoring; middleware that serves clean Markdown to AI crawlers.
A remote MCP server exposing six typed CRM tools so tasks can be captured by voice through Claude on mobile.
Clerk and Calendly webhooks; UTM-attributed lead capture from a profit-leak calculator and scorecard.
Technology
Next.js • Supabase • Clerk • Claude • MCP • Vercel