Answers. With records behind them.

Assistant

Ask, and see the records behind the answer.

By default Apple Intelligence answers on this Mac. Answers can show the records behind them, as a chart, a timeline, a map or cards you can open.

Answers without a model

• “Which apps used the most data today?”
• “Which apps opened the most connections today?”
• “What did Outbound block in the last hour?”
• “Which new servers did apps reach today?”
• “Which countries did apps connect to today?”
• “Why was (a site) blocked?” appears when today’s blocked records include a suitable domain.

Recognised suggestions read the records directly, even without Apple Intelligence. Change the wording to “this week” or “in the last hour” to change the ranking window.

Read and check an answer

Checked figures

Outbound holds the draft while it checks the figures (“Checking 12 figures against the records…”). It can try one correction, then show a records-based answer or an error. This checks the figures, not the truth of every sentence.

Supporting records

Highlights pick out up to three records. Charts compare figures or show a timeline; maps show countries. Cards open the matching app, destination or rule in the inspector when available.

Continue the conversation

Hover over an answer for “Copy answer” and “Ask again”. Up to four follow-up suggestions help you continue. Links in generated answers are removed before display.

Your coding agent

It answers on your existing subscription without an API key. Questions and the figures it reads go to its vendor or configured model provider under your sign-in. Outbound checks its figures against a record of the MCP tool results.

Who answered

Each answer names the model or agent that wrote it, such as “Answered by Claude Code”, or says “Answered by Outbound records”. If an agent quits mid-answer, its last message stays, such as a request to sign in again.

Which hosts Claude Code, Codex and Grok CLI contact on their own ›

Questions and history

New Chat

Open Assistant in the sidebar. Choose New Chat or press ⇧⌘N to start a conversation. Press Escape to stop an answer.

Dictate a question

With the input empty, choose Dictate a question. On macOS 26, dictation runs on-device and may download a missing language model before listening. On macOS 14 and 15, system speech recognition is not guaranteed to stay on the Mac.

Question length

Questions that need a model allow 1,600 UTF-8 bytes on the on-device model or 8,000 UTF-8 bytes with a coding agent. A character can use more than one byte.

Find and organise

Search conversations matches every typed word against titles and messages. Right-click a conversation for Rename… or Delete. Hide History and Show History fold the list away.

Stored on this Mac

The list beside the chat keeps conversations locally, with CloudKit syncing disabled. Edit › Undo can restore a deleted conversation while that edit remains undoable.

Earlier turns

An agent session can receive a summary of earlier conversation turns. Deleting a local conversation does not retract content already sent to the agent.

Explain

Explain one connection.

Select a connection and click Explain with Apple Intelligence. The model on your Mac says what the server most likely is, and whether this app usually connects there.

Example answer. The summary, Likely service and Usual for this app are generated advice. You can select and copy the text; it never changes a connection decision.

What Explain uses

Inspector facts

The app name and signing identity; host, port and transport; direction and decision; allowed and blocked counts; and the location label when known.

On-device processing

No packet data, URLs or history beyond those counts. Outbound sends no connection facts to Private Cloud Compute or a cloud provider for Explain.

Availability

You need macOS 26 or later with Apple Intelligence turned on. If it is unavailable, a hint explains what you can do where there is a fix.

Try Again

A failed explanation offers Try Again. Changing the selected connection facts clears the answer; click Explain with Apple Intelligence to request another.

In a connection alert

Open an alert’s details and click Explain this connection. With Apple Intelligence it runs on this Mac. With your coding agent or API-key provider, you see exactly what will be sent, with the website domain rather than the full host name, and nothing leaves until you click Send. Explaining never answers the alert.

Providers

Choose who answers.

Ask Assistant about this Mac’s network in plain words. Your own API key writes the Analytics overview and alert explanations.

Apple Intelligence. The default, and it runs on this Mac. Answers are written on your Mac, so nothing you see in Analytics leaves it. They show the records they used as charts, a timeline or a map, with follow-up questions to try. A Mac without Apple Intelligence shows a plain summary instead.

Your coding agent. Runs on the subscription you already pay for, with no API key. Choose Claude Code, Codex, Grok, Cursor or another agent installed on this Mac. It reads Outbound’s figures through Outbound’s own tools and answers under your own sign-in.

Your own API key. OpenAI, Anthropic, Google Gemini, OpenRouter, or a server you run, such as Ollama or LM Studio. Writes the Analytics overview and alert explanations with the model you choose. A server on this Mac can use http://localhost; any other server needs HTTPS. Your key stays in your keychain.

What your coding agent can and can’t do

Outbound talks to the agent over the Agent Client Protocol, the open standard coding agents use with editors. It finds the agents installed on this Mac, and you turn one on in Settings › Intelligence.

• Outbound gives it Outbound’s tools: your apps, the sites and countries they reach, what was blocked, and your rules.
• Outbound starts it in a folder of its own, allows only Outbound’s tools and refuses its requests for other work. The agent’s own settings still decide what it can do without asking, so this isn’t a sandbox.
• If it asks to change a rule, nothing changes until you approve it.
• Outbound tells you how to sign in when the agent needs it. If the agent doesn’t answer, the Analytics overview shows Outbound’s own summary.

Coding agents Outbound can answer with

• Claude Code
• Codex
• Grok
• Cursor
• Gemini CLI
• GitHub Copilot
• OpenCode
• Factory Droid
• Goose
• Kimi CLI
• Qwen Code
• Kilo
• Cline
• Auggie

And any other agent that supports the Agent Client Protocol.

Data choices

Choose what leaves your Mac.

Settings › Intelligence lets you choose who answers the Assistant and writes the Analytics overview. Explain in the inspector runs on this Mac with Apple Intelligence. Explain in a connection alert can also use your coding agent or API-key provider: you see exactly what will be sent, and nothing leaves until you click Send.

Keys and servers

App keys stay in the login Keychain. OpenAI-Compatible keys are saved per server address, so changing servers does not send the previous server’s key to the new one. Servers on localhost, 127.0.0.1 or ::1 need no key and can use plain HTTP; any other server needs HTTPS and a key. Outbound refuses a redirect to a different address or transport. Choose your own Model for a custom server.

Consent

Send figures to (provider) starts off. A new custom server address asks for consent again. API-key overview figures exclude addresses, full host names, file paths, rules and connection contents.

Preview and test

Show What’s Sent… shows the exact request for the range you’re viewing in Analytics, with your key hidden. If Analytics isn’t showing a range, it asks you to choose one first. Choose Test Connection to send a small “Say OK.” request. This test can run even while Send figures to (provider) is off.

Checked overview

Numeric values must match the supplied figures or list lengths. A failed draft falls back to Apple Intelligence or a plain summary. The check does not prove every sentence.

Agent setup

Answer with (agent) starts off. Claude Code and Codex need Node.js and can download pinned adapters through npx. Codex also starts in read-only mode; the agent’s own settings still govern what it can do without asking.

Agent profiles and tests

Claude Code shows Profile when more than one profile folder exists. Test (agent) starts a session asking for OK and waits up to 120 seconds. If sign-in is needed, Outbound tells you which command to run in Terminal.

Where your data goes

Stays on your Mac.

• Your rules and connection history.
• Answers from Apple Intelligence, and the check of their numbers.
• Server locations, from a database stored on the Mac.

Goes only where you point it.

• Your questions and the figures your coding agent reads, to the agent’s vendor or the model provider it is set up with, under your sign-in, once you turn it on.
• Figures to the model behind your API key, after you turn on Send figures to it. Show What’s Sent… shows the exact request for the range you’re viewing, and outboundctl ai ask --preview prints the exact request without sending it.
• An alert’s facts to your coding agent or API-key provider, only after you click Send under What will be sent.
• The same figures as Analytics, and your rules, to AI apps you add.

Never sent.

• The contents of any connection.
• To a model behind an API key: addresses, full host names, file paths or rules.
• Your figures, until you choose an AI and turn it on. A connection test sends a test request only.

The privacy policy has the details.

MCP tools

Let an AI app read the records.

The AI apps you connect see the same figures as Analytics and can read your rules. Nothing leaves through MCP by itself, and connection contents are never part of its results.

And 18 more, from traffic over time to the review queue, in the manual.

• Time ranges. Activity tools take hour, day, week, month or year; the default is day.
• Row limits. Ranking tools, blocked and new_apps accept 1–20 rows, defaulting to 10. rules accepts 1–200, defaulting to 50. network_summary takes only a window; firewall_status takes no parameters.
• Client setup. Add Outbound to merges the entry into the client’s existing configuration. Claude and VS Code need a restart afterwards. Codex’s TOML configuration is only appended to; existing lines are never rewritten or removed.

AI apps can query Outbound over MCP

Outbound includes a Model Context Protocol server, outboundctl mcp. Claude Code, Claude, Cursor, Codex, Gemini CLI and VS Code can ask about your apps, the sites and countries they reach, what was blocked and your rules. They see the same figures as Analytics. The contents of your connections aren’t part of it.

• Settings › Intelligence has an Add Outbound to button for each of them.
• An AI app can ask to add, remove, turn on or turn off a rule, switch profiles or block an app’s trackers. Nothing changes unless you allow it. For a rule, the approval panel says “Only you can approve this.” The AI app waits up to 180 seconds for your answer, and Edit › Undo takes an approved change back.
• You can switch rule changes off, so AI apps can only read. Or start the server with --read-only, which offers no tools that change anything.

What AI tools can see

By default AI tools see website domains, not full host names. Addresses show as “address”, local names as “local network” and app paths as app names. Your user name is never included. To share full host names, turn on Let AI apps see host names in Settings › Intelligence and start the server with --detail hosts. outboundctl mcp --print-disclosure prints the policy in effect.

AI approvals

Approve each requested rule change.

The on-device Assistant does not change rules. A coding agent answering in Assistant can request a change through MCP. The tools add_rule, remove_rule and set_rule_enabled file requests; Outbound’s background agent applies a change only after you allow it.

Ask Me Each Time or Off

Rule changes defaults to Ask Me Each Time, showing valid requests one at a time. Off keeps AI apps able to read but automatically refuses their rule-change requests.

Read the request

The panel names the AI app, shows the exact change and quotes its reason when supplied. Choose Don’t Allow (Escape) or Allow (Return). Invalid rules are rejected before they reach you.

Wait for the change

Allowing… stays visible until the change is applied. A change that loosens protection still gets the usual confirmation. Edit › Undo reverses an approved change.

Waiting requests

Each MCP server process allows three waiting rule changes. A fourth request is refused until you answer one.

Unanswered requests

The caller waits 180 seconds. If it withdraws an unclaimed request, a later click cannot apply it. The panel checks when you answer; it does not automatically close at the deadline.

Changes already claimed

If Outbound has claimed the request, the caller waits up to 120 seconds more. If that also expires, check Rules to see whether the change was applied.

Command line

Inspect a request before you send it.

outboundctl ai ask is a separate, read-only way to ask a provider about your network. It sends your question verbatim with the automatic figures, so anything you type in the question is sent too.

For ai ask, --window accepts hour, day, week, month or year and defaults to week. The CLI reads keys from environment variables; the app keeps its keys in Keychain.

Manual

Related in the manual

Keep exploring the Assistant, connection explanations, providers and AI apps.

• The Assistant ›
• Explaining a connection ›
• Providers and what is sent ›
• AI apps and MCP ›
• Approving AI changes ›
• Intelligence settings ›
• AI commands › · MCP commands ›