AI Activity
AI Activity
See every prompt, model request and tool call your AI agents make, what each one sent, and what Guard stopped, all stored on your Mac with secrets masked.
AI Activity shows what your AI tools and agents do: the prompts you send, the model requests they make, and the tool calls they run (shell commands, file reads and writes, MCP calls, web fetches), with what each one sent. Guard can stop secrets and risky tool calls before they happen. Everything is recorded on your Mac; nothing goes to Subcloak.
AI Activity is optional. It starts empty: connect a tool to fill it.
The four tabs
Section titled “The four tabs”| Tab | What it’s for |
|---|---|
| Activity | Sessions and their events, newest first, with the content of each event and what Guard did. This page. |
| Guard | Rules that redact secrets and block, ask or warn on tool calls. See Guard rules. |
| Usage | Tokens and estimated cost per day, project, session, model or agent. See Usage. |
| Connections | Each AI tool, how Watch sees it, and setup and undo. See Connect your AI tools. |
The status bar at the bottom shows each way Watch can see your tools, Hooks, Gateway, Telemetry and Network, with a dot and which tools it’s on for. Click one to open Connections. At the right, the lock shows how long content is kept, for example Content kept 7 days on this Mac; click it for the content settings.
How Watch sees your tools
Section titled “How Watch sees your tools”| Way | What it gives | Can it stop things? |
|---|---|---|
| Hooks | Prompts and every tool call with its input and output, as the agent runs them | Yes: block or ask before a tool call, block a prompt |
| Gateway | Every model request and response, and exact token usage | Yes: redact or block secrets in requests |
| Telemetry | Usage and tool metrics the agent reports | No |
| Network | Which app talks to which AI provider, how much and when, never the content | No |
Hooks, the gateway and telemetry are set up per tool on the Connections tab. The network extension covers every app at once. See Network extension.
Read the Activity tab
Section titled “Read the Activity tab”- Filter by agent and project, show only blocked or redacted, search
- Sessions by day
- What Guard did
- Allow it once, or always
- How long content is kept
- Sessions are grouped under Today, Yesterday, the weekday or the date. Each shows the agent, the project or first prompt, and Now while it’s active.
- The session header shows its time, model, new tokens and estimated cost.
- Events are listed in order, with Time, Event (Prompt, Model request, Bash, Read, Write, MCP call, WebFetch and so on), Command, path or prompt, Tokens · time, and a Guard label: Blocked, Redacted N, Asked, Warned, or Allowed after you allowed it.
- Filter with All agents, All projects, Blocked or redacted, and Search prompts, commands, files, which searches the stored content.
The event detail
Section titled “The event detail”Select an event to see what was sent.
| Part | What it shows |
|---|---|
| Header | The event, when, and what saw it (hook, gateway, telemetry or network) |
| Stats | For a model request: Model with the kind of login (Subscription login, API key or Gateway token), New and output tokens, Cached, Cost (an estimate; on a subscription, what the tokens would cost at API prices). For a tool call: Tool, Duration, Result (Succeeded, Failed or Not run). For a prompt: its length, Secrets found, and Guard. |
| Guard banner | What Guard did, for example 2 secrets redacted before sending or Blocked before it ran, the reason, each secret’s type, masked value and where it was, what the agent was told, and the rule, with buttons to allow it (see Guard rules). |
| Findings | Prompt injection in a tool result (MCP-5) or hidden text in an instructions file (AGT-6). |
| Content | For a model request, Sent (only what’s new in this request; earlier context folds away) and Reply. For a tool call, its input (Command, File, URL or Input) and Output. Secrets are highlighted and masked. Copy buttons copy the content with secrets still masked; file tools have Reveal in Finder. |
The model received [REDACTED: type] in place of each highlighted secret.
When there’s no content, the detail says why: content isn’t kept, telemetry reports usage but not content, the event was seen on the network only, or the content is older than your retention.
What Watch keeps
Section titled “What Watch keeps”Click the lock at the right of the status bar for the content settings.
| Setting | Default | What it does |
|---|---|---|
| Keep what agents send and receive | On | Keeps the content hooks and the gateway deliver: prompts, model requests, replies, tool input and output. Nothing is captured until you connect a tool. Turn it off to keep metadata only (tool names, timings and hosts); stored content is deleted at once. |
| Keep content for N days | 7 days | 1 to 30 days. Older content is deleted; the timeline keeps the event without it. |
| Projects | All on | Turn a project off and its content is never kept (and what’s stored is deleted); its events and Guard decisions still are. |
| Delete All Activity… | Deletes every event, all content and usage from this Mac, including older days. Guard rules and exceptions stay. |
- Secrets are masked before anything is stored.
- Every connect sheet says what Watch will keep, with a Keep prompts and content switch, so you can choose metadata only before you connect.
- The timeline itself (events, decisions and usage, without content) is kept for 30 days. Change both periods in Settings › Advanced (Keep AI Activity, Keep AI content).
- Everything lives in
~/Library/Application Support/Cloak Watch/AI/activity.sqlite, readable only by you.
While content is kept, Claude Code’s telemetry connection also asks Claude Code to include prompt text in what it reports to Watch (OTEL_LOG_USER_PROMPTS=1), along with tool details (OTEL_LOG_TOOL_DETAILS=1). With metadata only, Claude Code isn’t asked for prompt text. Those reports go to Watch on 127.0.0.1 only.
AI today
Section titled “AI today”Home and the menu bar show a one-line summary of today’s AI traffic, for example Claude Code → Anthropic 4.1 MB, 15 tool calls, 3 MCP servers. Click it to open AI Activity.