Your AI collaboration partner

Dear User,
how's the collaboration with your agent?

This year, you'll spend more hours with your AI agent than with most people in your life. Dear User was built to make the most of those hours.

$ claude mcp add --scope user dearuser -- npx -y @poisedhq/dearuser-mcp@latest Copy

Restart Claude Code afterwards so the tools appear.

Or see an example →
What you get

Dear User helps humans and AI agents understand each other better.

We read your setup like a new colleague who actually pays attention — then write back about where you're repeating yourself, where keys are leaking, and where your setup is quietly drifting apart.

Collab

How you two work together.

A score for your collaboration, your agent's "archetype", and concrete suggestions — not just "consider better instructions".

  • Catches patterns: corrections you make over and over
  • Suggestions with code snippets ready to paste
  • Tells you where your setup is strong
mcp__dearuser__collab
Security

Looks for keys you forgot.

Scans your Claude files for leaked API keys (OpenAI, Anthropic, GitHub, AWS, Stripe and 8 other types) and pulls signals from the platforms you already use.

  • 12 secret types caught in instructions, memory, hooks and skills
  • Pulls warnings from GitHub, Supabase, npm and Vercel — when you've set up tokens
  • Catches rules that contradict each other (e.g. "never force-push" + a hook that does)
mcp__dearuser__security
Health

What's quietly broken?

Draws a map of your setup and surfaces the silent issues: skills nothing calls, jobs writing to files nothing reads, rules that duplicate each other.

  • Finds "orphan jobs" — output that's never consumed
  • Catches overlap between skills, hooks and memory
  • Warns before your setup breaks silently
mcp__dearuser__health
Plus three more, included

Onboard

A guided conversation that helps you write your first CLAUDE.md or AGENTS.md.

Wrapped

Spotify Wrapped-style stats on how you and your agent actually work.

History

See the score trend over time and what improved since last scan.

"You've corrected Claude about response length four or five times this week — always 'keep it shorter'. Want me to remember it for you, so you don't have to say it again?"
From a collab letter
Get started

Three steps.

01 — Install

Add one line

Register Dear User in Claude Code. No accounts, no cloud.

claude mcp add --scope user dearuser -- npx -y @poisedhq/dearuser-mcp@latest

Restart Claude Code so the tools appear.

02 — Onboard

Answer a few questions

A short, guided chat so we know how you work. We tune the scoring and recommendations to your setup.

Run dearuser onboard
03 — Run first scan

Ask your agent

In any Claude Code session. The report opens locally at localhost:7700

Run dearuser collab
Honest FAQ

The questions people actually ask.

We're pre-launch. Here are the honest answers — including when the answer is "we don't know yet".

How is this different from just asking Claude to look at my setup?

Claude can definitely scan your files if you ask. The difference is memory and comparison: Dear User saves every scan locally so you can see the trend over time, and pulls real warnings from GitHub, Supabase and npm — not just pattern-matching in text. A conversation forgets. A database remembers.

Does my data leave my machine?

No. The scanners run 100% locally, and your database lives at ~/.dearuser/dearuser.db. The two exceptions require an explicit click from you: share report (uploads the report you choose, never raw data) and send feedback (sends the message you type). No background sync, no telemetry.

Why is it free? What's the catch?

No catch right now. We're pre-launch with close to zero users, and we want you to try it and tell us what doesn't work. A paid tier may come later (team features, monitoring, auto-scans) — but the core stays free and open source. Your data will never be the product.

Can Dear User change my setup without asking?

No. All recommendations sit as "pending" until you explicitly say yes via implement_recommendation. Dear User cannot write to your files, make commits, or push anything on its own. You always see a reviewable diff first.

What if my score is low? Should I worry?

No. First scan is often low — that's normal. The score isn't compared to an "ideal setup", it's compared to you over time. Low score just means there's low-hanging fruit to pick. Most patterns fix themselves after one solid CLAUDE.md or AGENTS.md update.

Does it work if I'm brand new to Claude Code?

Yes. The onboard flow is built for exactly that — a guided conversation that helps you write your first agent contract (CLAUDE.md or AGENTS.md). You don't need any existing setup to start. The more you build up, the more Dear User has to say — but day one is covered too.

How many users do you have?

Close to zero at launch. We built Dear User because we were spending 4+ hours a day with Claude Code ourselves and wanted to know how it was going. Expect rough edges in week 1. Feedback gets read personally — your input shapes the product, not a support queue.