Builds

WHOOP scores recovery and sleep, but the reports never see the rest of the day: meetings, workload, weather. And an assistant that calls the WHOOP API for every question burns a budget of 100 calls a minute shared by every user of the app.

A remote MCP server on Cloudflare Workers that connects WHOOP to Claude. Webhooks fill an edge database, so Claude answers from its own copy instantly without spending API quota, and an interactive dashboard renders inside the chat. Invite-only, for up to 100 members.

Letting an assistant write to your CRM is a gamble when approval lives in a client setting someone can flip to "always allow". HubSpot's own agent tools have no undo, so one bad bulk write is permanent.

A standalone MCP server for HubSpot with 79 CRM tools and a 7-tool safety layer. Every write produces a preview, nothing reaches HubSpot until a person approves it, and each approved change records an undo snapshot and an audit entry. It runs in Claude Cowork, where the people who own the CRM work, not only in a developer tool.

Ops teams spend hours copying deals, contacts and tickets out of HubSpot into spreadsheets and chat threads, cleaning them up, then pasting the results back. Bulk cleanup in the HubSpot UI is a clicking marathon, and a raw API script has no preview and no undo.

An open-source Claude Code plugin that runs HubSpot in plain English. 44 specialist sub-agents handle contacts, deals, workflows and hygiene. Every write is previewed and approved, destructive changes need the exact record count, and long cleanups run as loops that pick up where they left off.

Teams starting with agents pick a pattern because it's popular, not because it fits. A single prompt gets asked to do what needs a pipeline, or a multi-agent setup gets built for a job one skill could handle. The cost shows up weeks later as rework.

A Claude Code plugin and skill that matches the system you describe against the live Encyclopedia of Agentic Coding Patterns and returns the 4 to 8 that actually apply, out of about 300, with the reason each one fires. It leads with the gaps: the destructive path with no approval policy, the router with no eval gate.

Most prompts for recurring agent jobs are numbered procedures with no way for the agent to check its own work. They run once, then quietly drift, and someone ends up inspecting every output by hand.

A Claude Code skill that interviews you before you write the prompt, then drafts one built on the outcome, the guardrails and a self-check, and recommends how to run it: a single agent, a workflow, a loop or a scheduled routine. It drafts only and never schedules anything for you.

Teams shipping LLM features keep prompts in code, spreadsheets and chat threads, with no version history, no record of what each call cost, and no way to tell whether a prompt change made things better or worse. The usual answer is another hosted observability vendor.

A lightweight, self-hosted prompt registry. Every prompt change is a Git commit with full history and rollback, every LLM call logs model, tokens, latency and cost, and evaluations and A/B tests show which version wins. It also scans prompts for PII and API keys. Prompt content lives in Git, not a database, and it ships on npm with an optional web dashboard.

When an agent does something unexpected, logs tell you what happened, not why. Finding the turn where it went wrong means digging through raw transcripts, token counts and tool errors by hand.

The investigation layer for OpenClaw. It captures every session, turn, LLM call and tool execution, then lets you replay a conversation turn by turn with cost, token and timing notes, check every agent's health at a glance, and query across sessions for cost patterns, tool failures and slow calls.