Introducing Synthetic Trust

A local audit trail for Claude’s tool activity and MCP connector changes

xCLAUDE records and classifies tool activity from Claude Code and wrapped MCP servers — locally on your Mac, in a record you control.

Download for Mac

Free · Audited locally · No account · No telemetry

macOS 13+ on Apple Silicon

Windows coming soon

xCLAUDE Gateway

BETA

Audited locally · No account · No telemetry

SourcesDetectionsMCP ChangesClaude Code
1628
TOTAL
9
LOW
14
MEDIUM
3
HIGH
5
CRITICAL
TIMESEVERITYSOURCETOOL
17 Jun, 17:22:30NONEGitHubget_file_contents
17 Jun, 17:12:57NONEClaude CodeRead
17 Jun, 17:04:19MEDIUMNotionnotion-update-page
17 Jun, 16:37:10NONEGitHubget_file_contents
17 Jun, 16:37:02NONExcg-filesystemread_multiple_files
17 Jun, 16:25:06NONEClaude CodeBash

Why people install it

Every other tool you use has a log. Your AI shouldn't be different.

01

Keep a trail you can come back to

Tool activity from Claude Code and wrapped MCP servers stays in a local, reviewable record you control — across sessions and over time.

02

Know what deserves a second look

xCLAUDE classifies activity as it's recorded, surfacing credentials, PII, prompt-injection patterns, data exports, email actions, and changes to what MCP connectors advertise or where they sign in.

03

One audit trail across Claude Code and Desktop MCP

Bring Claude Code tool activity and wrapped MCP traffic from Claude Desktop into one place, so you can review both without piecing together separate logs.

How it works

One download. One audit trail. Keep using Claude as usual.

01

Download

Download xCLAUDE for macOS and install it like any other Mac app.

›xclaude.ai/download
02

Connect

Enable Claude Code auditing and connect or wrap the MCP servers you want xCLAUDE to observe.

›claude_desktop_config.json
03

Review

Tool activity is recorded locally, classified by xCLAUDE, and available to review whenever you need it.

›ls ~/Library/Application\ Support/xCLAUDE\ Gateway/wrappers/

The dashboard

Review tool activity, one line at a time.

A native macOS app with a menu bar icon. It refreshes every 2 seconds and shows each event's severity, source, and tool call details. No telemetry leaves your Mac.

Severity and source for every event
Tool name and arguments for each call
Connector changes and sign-ins in MCP Changes
xCLAUDE Gateway — Detections viewxCLAUDE Gateway — MCP Changes view with a flagged change openxCLAUDE Gateway — Detection detailxCLAUDE Gateway — Claude Code viewxCLAUDE Gateway — OAuth sign-in recorded in MCP ChangesxCLAUDE Gateway — Remote connector detailsxCLAUDE Gateway — Claude Code sourcexCLAUDE Gateway — Add source modal

Connectors

Audit the MCP tools you already use.

xCLAUDE includes a catalog of common MCP connectors. Connect them through xCLAUDE instead of Claude Desktop’s native Connectors, and their MCP activity is recorded and classified in your local audit trail. Six connect in one click; Slack and Google services use a guided setup first.

Notion
Gmail
Slack
Linear
Atlassian
GitHub
Drive
Calendar
Stripe
Apollo

Native Claude Desktop Connectors (Settings → Connectors) bypass xCLAUDE entirely.

See how to reconnect through xCLAUDE →

Claude Code

Your coding agent, on the record.

xCLAUDE keeps an independent, local audit trail of Claude Code’s tool activity. Install the hook from Add source and it records tool calls after they finish, including built-in and MCP tools. It also records session starts and ends and MCP elicitations, and checks the activity against the detections supported for Claude Code.

Review activity by session or project, filter by severity or tool, search the trail, and export it when needed. You can also wrap Claude Code’s MCP servers for protocol-level visibility.

  • One-click setup from Add source
  • Flags writes to xCLAUDE’s own audit data
  • Stored in the same local trail as connector activity
xCLAUDE Gateway — Claude Code view with an open detection

Coverage

Tool activity, classified and logged.

xCLAUDE records the tool activity it covers and classifies it across severity levels, so you can review what deserves attention without blocking the underlying call.

CRITICALHIGHMEDIUMLOWNONE

Transparent about what it does and doesn't cover. See the full breakdown →

See what it doesn't cover →

EventTriggers when →

Prompt injection

CRITICAL

Tool arguments or response contain instruction-override or jailbreak patterns.

Credential detected

CRITICAL

Tool arguments or responses contain a known API key or token format.

Audit trail modification

HIGH

A Claude Code tool call writes to or deletes data from xCLAUDE’s own data folder.

Email send warning

HIGH / MEDIUM

Tool arguments or responses contain email-sending language, or the tool name signals sending, replying, or composing.

Connector changes

HIGH / MEDIUM

Shown in MCP Changes when what a connector advertises changes, or when its sign-in changes (authorization server or permissions). Severity comes from matching rules, not the change itself.

Protocol tripwire

HIGH / MEDIUM / LOW

A server asks the client for something, such as a form or sampling request, uses an unknown method or protocol version, or returns input_required. HIGH when a Claude Code form field asks for a secret.

Structured PII

MEDIUM

Tool arguments or responses contain well-formed personal data such as email addresses, international phone numbers, IBANs, credit cards, passports, or national IDs from several countries.

Data export warning

MEDIUM

Tool arguments contain bulk-extraction language — export, download, or dump targeting data or databases — or a tool response contains export language pointing at an explicit destination.

PII detected

LOW

On wrapped and remote connectors, tool arguments contain named entities: person names, organizations, or locations (async, best-effort).

Tool call allowed

NONE

Routine calls with no findings. They are still logged so the trail includes ordinary activity, not just incidents.

Questions

Things people ask before installing.

No. xCLAUDE is an independent audit layer built on top of the Model Context Protocol, which is an open standard. Claude and Claude Desktop are trademarks of Anthropic.
Nothing goes to us — there is no xCLAUDE server, no account, and no telemetry. The audit log, dashboard, and detection engine run entirely on your Mac. Connector traffic goes from your Mac directly to each provider; xCLAUDE observes and records it locally.
No. You can use Claude Desktop, Claude Code, or both. With Claude Desktop, xCLAUDE audits the MCP connectors you route through it. With Claude Code, a hook records tool calls after they finish, plus session starts and ends. The two work independently.
No — and it's a limit of where the traffic goes, not a missing feature. xCLAUDE audits what passes through your Mac. When you use Claude at claude.ai, it reaches your connected tools from Anthropic's servers, so that traffic never touches your machine and there is nothing for xCLAUDE to observe. Claude Desktop and Claude Code run on your machine, which is what makes the audit possible — though Claude Desktop's own native Connectors also bypass xCLAUDE, so add your services from the Sources tab instead. If you use the browser today, install Claude Desktop or Claude Code and route your tools through xCLAUDE there.
Yes. Install the hook from Add source. Claude Code tool calls are recorded after they finish and classified in their own view, along with session starts and ends and MCP elicitations. You can also wrap Claude Code's MCP servers to see what happens at the protocol level: what each server advertises, its stderr, and server-initiated requests.
No. Native Claude Desktop Connectors (Settings → Connectors) connect directly to provider servers, bypassing xCLAUDE entirely. xCLAUDE audits the MCP servers configured through it — including the connectors you add in the app.
Yes — but it's guided, and the key is yours. xCLAUDE ships with no Google credentials of its own: you create your own project in Google Cloud and paste your own OAuth client into the app. A 4-step wizard with deep links walks you through it — Cloud project, OAuth client, Preview enrollment, paste your credentials. One free OAuth client serves Gmail, Calendar and Drive, and your credentials are stored in the macOS Keychain. Two things to know before you start: Google's Preview enrollment form requires an email on a custom domain (the Google account you connect and audit can be a regular Gmail), and approval arrives by email, usually within a couple of days. Because the client is yours, Google shows a "Google hasn't verified this app" screen when you authorize; that's expected — you continue past it. Choosing Publish app on the consent screen avoids the 7-day expiry. If you leave the project in testing, the authorization expires every 7 days.
No. Everything runs on your Mac. The proxy forwards each message as it arrives and analyzes a copy, so detection doesn't delay the call. Claude Code events are recorded after each tool call finishes.
Before removing the app, open xCLAUDE, go to Settings, and click Uninstall to restore your MCP configuration. If you installed the Claude Code hook, remove it first from the Claude Code source (Uninstall). Then quit xCLAUDE from the menu bar, drag the app to the Trash, and restart Claude Desktop.
Not yet. The current beta requires macOS 13 or later on Apple Silicon (M1 or newer).
Yes. xCLAUDE is MIT-licensed and the source code is on GitHub.

Make Claude’s tool activity reviewable.

Stay in the loop. Get notified about new detections, releases, and product updates.