Built on Claude.

Not every task needs the same model. We route based on complexity and cost:

Task	Model	Why
Intent classification, tagging, simple transforms	Claude Haiku	Sub-second latency, fractions of a cent
Complex reasoning, report synthesis, multi-step analysis	Claude Sonnet	Best reasoning-to-cost ratio for production workloads
Architecture decisions, ambiguous judgment calls	Claude Opus	Deep reasoning when the stakes justify the cost

This tiered approach lets us serve clients at price points ($297 audits, $997 sprints) that would be impossible with a single-model architecture. The routing logic runs in our 18-layer agent pipeline and makes the model decision per-task, not per-session.

Our internal production platform — the DYOE Agent Stack — is an 18-layer multi-agent pipeline that coordinates Claude agents for autonomous execution:

• Tool use — agents call tools (web search, file operations, API calls) within bounded permissions
• Persistent memory via Mem0 — cross-session memory that survives restarts and carries client context across interactions
• DeerFlow research integration — multi-agent research framework that powers deep competitor analysis, market scans, and revenue audits
• Hard-interrupt safe word — any human operator can halt the pipeline instantly; the system is designed to be interrupted at any point
• Structured outputs — agents return typed JSON responses for downstream consumption, not free-text blobs

Every DYOE Way deliverable — every audit, every campaign, every SplitLedger edge function — is built on top of this stack. We don't sell the stack itself. We sell the outcomes it produces.

We use Model Context Protocol (MCP) to connect Claude to external systems safely. MCP servers give Claude structured, permissioned access to data and APIs without hardcoding credentials into prompts or relying on screen-scraping.

Our production MCP integrations include:

• Airtable — CRM for leads, clients, and revenue tracking
• n8n — workflow automation (form submissions, alert routing, campaign triggers)
• DeerFlow — research agent orchestration for the AI Business Audit skill
• GitHub — code review, PR management, deployment triggers
• Supabase — database management and edge function deployment for SplitLedger

Each MCP server runs as a registered integration in Claude Code. This means we can build, ship, and iterate on client workflows without leaving the Claude environment.

Our primary development tool is Claude Code — Anthropic's official CLI for Claude. Everything from website deployment to edge function development to client outreach automation is built and iterated within Claude Code sessions.

We use Claude Code for:

• Writing and deploying Supabase edge functions (SplitLedger's parse-statement, budget-advisor, dashboard-insights)
• Building and iterating on website pages (this very page was built via Claude Code)
• Managing GitHub repos, PRs, and deployments via CLI agents
• Creating and testing n8n workflows
• Authoring Mailchimp campaigns with segmented personalization

Claude Code isn't a toy we use for demos. It's the production development environment for every DYOE Way venture.

We use the Anthropic API's prompt caching feature to reduce latency and cost on repeated context windows. When a client's audit runs, the system prompt and reference material are cached for the session — subsequent agent calls within the same run hit the cache, not a cold prompt.

Combined with model routing (Haiku for simple tasks, Sonnet for reasoning), this keeps per-client delivery cost low enough to offer fixed pricing at $297 and $997 without margin compression.

Building with Claude means building within Anthropic's safety framework. We follow these principles:

• Human in the loop, always. No autonomous campaign sends, no unreviewed deliverables, no unsupervised client-facing outputs. Every output is reviewed by DYOE Way before it ships.
• No invented proof. We don't generate fake testimonials, fabricated metrics, or theoretical outcomes presented as real results. Our case studies use real numbers from real work.
• Bounded permissions. Agents operate within declared tool-use boundaries. A research agent can't send an email. A campaign agent can't access financial data.
• Aligned with Anthropic's Acceptable Use Policy. Every use case we build and sell is explicitly within the Anthropic AUP. We don't build deceptive agents, surveillance tools, or manipulation systems.
• Data handling. Client data stays in our controlled infrastructure (Supabase, Airtable). We do not train models on client data. We do not share client data with third parties.

When we say "Powered by Claude" we don't mean we use the chatbot. We mean every layer of every product is calling the Claude API in production, with intentional model routing, prompt caching, tool use, structured outputs, and persistent memory. Claude is not a feature we bolted on — it's the operating system our business runs on.