v1.0.0-beta.5

What's new

Claude Code hook integrity. If the capture hook disappears from ~/.claude/settings.json without an in-app Uninstall — another program, or the agent editing its own settings — xCLAUDE now notices. You get a dismissible notice with one-click Reinstall, and an app.cchook_removed marker is written to the audit trail, so the removal stays on the record even after the notice is dismissed. The check runs at app start and every 15 seconds, so removals that happened while the app was closed are caught too.

Wider Google scopes. Gmail moves to gmail.modify and Calendar to calendar.events, so drafts, labels and event changes are audited end to end. Existing Gmail and Calendar connections keep the scopes they were granted at connect time — to pick up the new ones, remove the connector in xCLAUDE and connect it again.

Google setup wizard. The consent-screen step now recommends publishing your Cloud app instead of adding test users: you authorize once, rather than reconnecting every 7 days. The unverified-app warning still appears; you continue through Advanced.

Fixes

The window's drag strip no longer collapses under the macOS traffic lights when a notice is shown. List heights in Detections and Claude Code are now measured from their container instead of computed from fixed constants.

Documentation

• Claude Code is presented as a first-class audited source: the tagline, requirements and quickstart no longer assume Claude Desktop.
• Scope and limits are consolidated into one section with a coverage table of nine surfaces, including what xCLAUDE cannot see and why.
• The in-app About names Claude Code, and the trademark notice covers it.

Requirements: macOS 13 or later, Apple Silicon.

v1.0.0-beta.4

The biggest release since going public. Beta.4 brings Claude Code auditing — a full audit source with its own view — plus credential masking, a sharper detection engine, and substantial reliability work across the audit pipeline.

Claude Code auditing (new audit source)

• Native session auditing via a session hook: every tool call in a Claude Code session — native tools and MCP tools alike — is captured, classified and recorded, with the same detection engine as wrapped traffic. Install/uninstall from Add source in one click.
• New Claude Code tab: the claude-code slice of the trail as its own view — severity cards, faceted filters (Tool, Session, Status, Project), free-text search, session separators, request↔response outcome, and filtered export.
• Wrapping Claude Code's MCP servers (documented in the README): adds what only the protocol level shows — the server's tool manifest, stderr, and server-initiated traffic. Calls seen by both sources are correlated by tool-use ID: paired rows carry a link indicator, and CSV export gains cc_tool_use_id and paired_source columns.

Credential masking

• Detected credentials are now masked before the event is written: a 10-character prefix plus an irreversible HMAC-SHA256 fingerprint, keyed by a per-install salt. The same salt keys both audit sources, so one credential carries the same fingerprint wherever it appears. The one deliberate exception to capture-all — everything else is recorded as captured.

Detection engine

• tool_manifest_changed is now graded: new tool surface or injection markers in descriptions flag HIGH; docs/typing rewrites and add/remove flag MEDIUM — less alert fatigue on routine vendor updates.
• Inbound data_export_warning now requires an explicit destination, eliminating a class of false positives on ordinary document content.
• Result-text deduplication in the frame processor — no more doubled findings from mirrored content parts.

Filters & UI

• New date-range picker for the Custom time filter — one calendar for both ends of the range, full keyboard support, no new dependencies.
• Correctness fix: custom date ranges now select local calendar days, including DST transition days. Previously the bounds were UTC days, which could shift the range by your timezone offset.
• Free-text search and custom ranges in both views; restyled filter dropdowns; inline Clear filters; honest tool labels on enrichment rows.

Reliability & performance

• OAuth token refresh is serialized across processes, preventing server-side grant revocations (the cause of recurring Notion re-logins).
• Session log compaction: finished sessions consolidate into per-day files, keeping the reader fast as the trail grows. Nothing is deleted — events are preserved verbatim, with crash-safety guarantees that fail toward duplication, never loss.
• DMG notarization and stapling now run inside the build; CI on every push.

Checksums

SHA-256 (xCLAUDE Gateway-1.0.0-beta.4-arm64.dmg):
0639590b31331997289d4ee7aa2b5622f30463156fb60d674e508a92c9e56265

Requires macOS 13+ on Apple Silicon.

v1.0.0-beta.3

Reliability release. Fixes a token-rotation race between concurrent proxy processes that forced daily re-logins on connectors with rotating refresh tokens (e.g. Notion): the losing process no longer deletes freshly rotated credentials and recovers automatically. OAuth failures in the audit log now carry failure-mode context, and unreadable Keychain blobs are treated as absent credentials instead of erroring on every request.

SHA-256 (DMG):
141a1c2f3606b534caeab53cc856ed5a3fec7a220435bfe4aabb6db30c2dd0f0

v1.0.0-beta.2

Maintenance release with fixes from a systematic QA pass: stale re-login alerts after re-adding a connector, leftover-credential notices on failed removes, out-of-band config changes now detected within seconds (with a re-add prompt), restored the Add connector button in the empty state, plus input-validation hardening across the IPC surface.

SHA-256 (DMG):
49a75db01d29c4165527720be69fbb31617ac548fd87519fabfb4df7f3766b3f

xCLAUDE Gateway 1.0.0-beta.1

xCLAUDE Gateway 1.0.0-beta.1

First public release. xCLAUDE Gateway is a local audit layer for Claude
Desktop's MCP traffic: it sits between Claude Desktop and the tools it talks
to, records every call, classifies what it sees, and warns you — it never
blocks. Your traffic still reaches the service it's addressed to.

What it does

• Records every JSON-RPC frame between Claude Desktop and your
  connectors — both directions, requests and responses — to per-session
  logs on your Mac.
• Classifies what it sees across 7 risk categories — credentials,
  prompt injection, email-send warnings, data-export warnings, named-entity
  PII, checksum-validated structured PII, and tool-manifest changes — and
  4 severity levels, plus a baseline entry for every normal call.
• Warns, never blocks. A Detections view with severity, category and
  time-range filters, JSONL/CSV export, a menu-bar counter of flagged calls,
  and re-login alerts. No traffic is rerouted or withheld.
• Local-first. Everything is audited on your Mac: no account, no
  telemetry, no analytics. Credentials and tokens stay in the macOS Keychain.

Connect your services

• One-click connectors: Notion, Linear, Atlassian, GitHub, Stripe, Apollo.
• Google (Gmail, Calendar, Drive): a guided 4-step in-app wizard — Cloud
  project, OAuth client, Preview enrollment, paste your credentials. No
  Terminal involved.
• Slack: a 2-step wizard — one click creates your own Slack app from a
  preloaded manifest, then paste your Client ID.
• Local MCP servers you already run are wrapped by Install in one click —
  and reverted just as easily, with a backup of your original config.

Beta

xCLAUDE Gateway is in beta — expect rough edges. Check for updates from the
Settings drawer, and use "Request a connector" for anything you're missing.

---

Requirements: macOS 13 or later, on Apple Silicon (M1 or newer). The
current build is arm64-only; Intel Macs are not supported.

Website: https://xclaude.ai

xCLAUDE Gateway records, classifies and warns; it never blocks.