Your AI agents act in your name
They use your email, your keys and your accounts, mostly out of sight. Excellent is a free Mac app that shows what they did. Turn it on, and it blocks or asks about the risky actions it checks for. Unless you choose to pair it with our team, nothing about you or your agents is sent to us.
If your agents sent you a notice every time they used your name, your Mac would look like this.
jordan@mac portfolio-site % agent "Find reference images for the new portfolio" • Starting 9 helper agents image lead · color lead · type lead · 6 more • Collected 214 reference images from 6 archives • Wrote board.html, palette.json and sources.md ✓ Done. Your reference board is in ~/Sites/portfolio-site/refs jordan@mac portfolio-site %
Seen by you: 0
Illustrative: a made-up user, Jordan. Each kind of event was found in real agent logs on one Mac; the Excellent cards use the app's own words.
Cards marked Reported are real incidents. Sources:
- April 2026: a Cursor agent running Claude Opus 4.6 deleted PocketOS's production database and its volume-level backups in one call to Railway's API; Railway restored the data. The Register, Apr 27, 2026
- July 2025: Replit Agent deleted data from the database of SaaStr co-founder Jason Lemkin's app; he restored it with Replit's Rollback. Replit, Jul 29, 2025
- Reported March 2026: Claude Code ran a Terraform destroy that wiped Alexey Grigorev's setup for DataTalks.Club and AI Shipping Labs, including a database with 2.5 years of records and its snapshots; Amazon's support restored the data in about a day. Tom's Hardware, Mar 7, 2026
- January 2026, a user's report: asked to compare JSON files and change no other file, Gemini CLI deleted the *_new.json files without asking. GitHub, google-gemini/gemini-cli #15822, Jan 2, 2026 (a user's report)
- July 2026, a user's report: in auto mode, a Claude Code session asked to run a project with Docker deleted the contents of the home directory; closed as stale with no maintainer reply. GitHub, anthropics/claude-code #74539, Jul 5, 2026 (a user's report)
Reads the logs of
Claude Code
Codex
Grok Build
GitHub Copilot CLI
Gemini CLI
Qwen Code
Trademarks belong to their owners; no endorsement implied.
Excellent is a free Mac menu bar app that shows what your AI agents did with your email, your keys and your accounts
01 Overview
It reads the logs your agents already keep
Your agents keep a diary of what they did, helpers included. Excellent reads those of Claude Code, Codex, Grok Build, Copilot CLI, Gemini CLI and Qwen Code, on your Mac. A first scan takes seconds.
02 Findings
See what was done in your name
Your email sent to websites or typed into forms. Keys sitting in logs. Secrets files read. History overwritten. Checks skipped. Each one says which agent and where.
03 What you can do
Fix some of it with a button
Have your own Claude Code or Codex take your email out of scripts, three files at a time. Clear keys from logs. Repair hooks whose Node version is gone. Excellent backs up a settings file before it changes it. Clearing keys from logs can't be undone.
04 Rules
Give Claude Code and Codex the same rules
A rule written for one often never reaches the other. Excellent shows which rules reach only one, and on your OK adds them to a file the other reads.
05 Notifications
Stop it before it happens
Once you turn it on, Excellent checks each action before it runs, in 9 agents. It blocks your email on its way to a website and asks before a force push or a look at your keys. It can pause an agent until you decide.
06 Privacy
Unless you choose to pair it with our team, nothing about you or your agents is sent to us
No account. No analytics. Apart from a daily check for updates, which you can turn off, Excellent goes online only for things you start: Check GitHub and GitLab, Protect Main, and the Servers preview. Once on, the preview checks your servers every 5 minutes. Explain and agent fixes run your own agent, which talks to its provider.
07
Make your rules hold, whatever the wording
Excellent can turn your rules into the Mac's own sandbox for Claude Code and Codex. The Mac then refuses to open or change the files and folders you closed, however a command is written. Undo takes it out.
Here's the app, with a made-up user's logs
Click the star, the rows and the sidebar. Your own scan stays on your Mac.
Get the Mac appWhat your AI agents did on this Mac, from their own logs.
In the app, incident rows name the agent. Here they say “Agent”.
Not ready for an app? Ask your agent to check
Paste one prompt into Claude Code or Codex. Your agent saves a small script, checks its fingerprint, tells you what it does, and runs it. The script reads your agent logs and writes one page on your Mac. Your agent is told not to read the logs itself, so what's in them stays on your Mac. Only the counts and the report's location pass through it.

Needs Python 3.9 or later. Reads only. Makes no network calls. Writes one file: ~/excellent-free-scan/report.html.
SHA-256 of the scriptfdca306b2d1cc281ce2bf90c70846a9995916e76b90201f0cbe50ff2946a32c7

- 6requests carried my email to 3 websites
- 2API keys in plain-text logs
- 9secrets files opened
- 1force push to main
- 3rules reach only one agent
- 0safety checks failing
Get Excellent, free for your Mac
Excellent for Mac
- Free. No account.
- For Macs with Apple silicon, macOS 14 or later.
After you open it, press Protect My Agents. Excellent shows you what it found and what it will change before it changes anything.
One email when Excellent for Mac is ready to download. Nothing else.
Questions: help@excellent.so
Pro, in January
Questions
What does Excellent do?
It reads the logs of Claude Code, Codex and four other agents on your Mac and shows what they did in your name: your email sent to websites, keys in logs, secrets files read, history overwritten, rules that never reached the agent. It can fix some of it: clear keys from logs, take your email out of scripts, share rules and repair broken hooks. For the rest, a button turns on a check for next time. Once you turn it on, it checks each action before it runs, in 9 agents.
Does anything leave my Mac?
Unless you choose to pair it with our team, nothing about you or your agents is sent to us. Excellent reads your agents' logs on your Mac and keeps what it finds there. No account, no analytics. It checks for updates about once a day, which you can turn off. A few things leave when you start them, such as Explain, agent fixes, Check GitHub and GitLab, Protect Main and the Servers preview. Explain hands a masked finding to your own Claude Code or Codex, on your plan. An agent fix gives it the file's path, and it reads and edits that file itself. Replace this key opens the provider's key page in your browser. You make the new key there. The check's replies, which can quote your rules, reach your agent's provider.
Which agents does it work with?
It reads the logs of Claude Code, Codex, Grok Build, GitHub Copilot CLI, Gemini CLI and Qwen Code. Its live check runs in those and in Cursor, Factory Droid and Devin. It recognizes 43 agents in all.
What does it change on my Mac?
Nothing until you say so. Before you press Protect My Agents, it lists what it will change. It adds one check to each agent's settings, backing up each file first. It also tells Codex to trust it and starts at login. Turn it off and it takes its check back out of each agent's settings, and changes nothing else. Codex trust, shared rules, a sandbox and commit watching stay until you undo each. It never opens your Documents, Desktop, Downloads or iCloud folders.
Is it free?
Yes. Everything that protects this Mac is free. Pro comes in January, for people who want more than one Mac or their servers in one view, alerts away from the desk, and a weekly digest. It will be $6 a month or $60 a year, and $40 for the first year if you join the list by 11:59 p.m. PT on Monday, Nov 30.
Can an agent get around it?
Its live check reads what a command says. An agent set on getting around it sometimes can. For the file and folder rules that matter most, Excellent can turn them into Claude Code's and Codex's macOS sandbox, if you choose. The sandbox doesn't care how a command is worded.



