AI ACTIVITY MAP

What your AI went through
and what it touched, at a glance.

See the work you handed to Claude Code as one flow per request (Claude Code → MCP → read outside sites → ran commands → wrote files) and as a connections map: AI → MCP server → tool → touched. Viewing it is free, and the records stay on this PC.

We don't stop your AI. We only brake when something is risky.

SCENES

When it protects you.

01

You want to check later what Claude Code did while you left it working

Each request becomes one line: which MCP it used, whether it read outside sites, ran commands and wrote files. Click one to see that work's connections map and timeline.

02

You want to know whether an unknown MCP server slipped in

Each MCP server gets a label: Started by your AI, or Started from outside / unknown. Outside or unknown ones are drawn with a dashed line and ⚠, so they stand out in the map too.

03

Did it run a command right after reading outside text?

The classic prompt-injection pattern is: read outside text, then run or send what it says. NEXORIA counts boundary crossings such as Outside site → local file first, and shows each site read by its landing host and the number of instruction-like lines.

Work stories: for each request, the MCP servers the AI went through and what it did (read outside sites, ran commands, wrote files) on one line. The screen is an example with made-up data (Japanese screen).
Work stories: for each request, the MCP servers the AI went through and what it did (read outside sites, ran commands, wrote files) on one line. The screen is an example with made-up data (Japanese screen).
Connections: AI → MCP server → tool → touched. Line width is the count, color is the judgement (safe, needs a look, danger), and dashed lines with ⚠ mark MCP servers started from outside / unknown. Example with made-up data (Japanese screen).
Connections: AI → MCP server → tool → touched. Line width is the count, color is the judgement (safe, needs a look, danger), and dashed lines with ⚠ mark MCP servers started from outside / unknown. Example with made-up data (Japanese screen).

WHAT HAPPENS

What happens
when you use it.

  1. 1Ask Claude CodeHand off work as usual
  2. 2AI Brake records itEach time it uses commands, writes, MCP and similar tools, only the kind and judgement are kept
  3. 3Open AI ActivityThe map is built on the spot from today's or the last 7 days' records
  4. 4Click what catches your eyeClick a node to list only the calls that went through it

MECHANISM

How it works.

What you see

Boundary crossings (outside site → local file, secret → sent out, outside / unknown → MCP, writes / deletes outside the work folder), a story for each request, the connections map, recent calls and a timeline. ▲ marks where secrets, outside text and sending out all came together.

Where the data comes from, and which AI

It uses the records AI Brake keeps through Claude Code's hooks, plus PC Watch's records of MCP servers starting. Right now it records Claude Code with the NEXORIA hook installed.

What it never shows

Paths, file names, URLs and command contents are never shown (only kinds and counts). For outside sites it keeps only the landing host and a summary (language, length, number of instruction-like lines). Excerpts of text read are kept for 14 days only if you turn that setting on (off by default).

PRIVACY

What goes out,
what stays in.

What is sent out

  • Nothing. The records and the map stay on this PC

What is never sent

  • Paths, file names, URLs and command contents
  • Excerpts of text read (kept on this PC for 14 days only if you turn the setting on)

HONEST

What it can do,
and what it can't.

What it can do

  • See, per request, the MCP servers the AI went through and what it did
  • See AI → MCP server → tool → touched as a map
  • Mark MCP servers started from outside / unknown
  • See how many outside sites were read, their landing hosts and instruction-like lines
  • Viewing AI Activity is free

What it can't do / limits

  • This screen is for looking. Stopping actions is AI Brake's job
  • Only work by Claude Code with the NEXORIA hook installed is recorded
  • Telling who started an MCP server needs PC Watch records (starting PC Watch needs a plan or the free trial; without records it shows No start record)

PRICING

Free for 10 days from sign-up.

After that, Pro is ¥980/month (1 PC) and Small Team is ¥2,980/month (up to 5 PCs). Your decision history and settings stay available even after your plan ends.

See pricing details→

QUESTIONS

Frequently asked questions

Where do I find it?+

In the app, open AI guard → AI activity on the left. It is built from the records when you open the screen, so nothing keeps running in the background.

Does it cost anything?+

Viewing AI Activity is free. Starting PC Watch, which is used to tell who started an MCP server, needs a plan or the free trial.

Does it record what the work contained?+

No. Only the kind of action, how many times, and the judgement. Paths, file names, URLs and command contents never appear on screen or in the records.

DOWNLOAD

Start by installing the Windows version.

The first 10 days from sign-up are free. You can check the installer's SHA-256 (NX2026.10.22.v37 / approx. 61.7 MB) in the download section.