<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0"><channel><title><![CDATA[OffiAI blog]]></title><description><![CDATA[OffiAI blog]]></description><link>https://offiai.hashnode.dev</link><generator>RSS for Node</generator><lastBuildDate>Wed, 18 Nov 2026 14:57:04 GMT</lastBuildDate><atom:link href="https://offiai.hashnode.dev/rss.xml" rel="self" type="application/rss+xml"/><language><![CDATA[en]]></language><ttl>60</ttl><item><title><![CDATA[What four coding agents expose on your Mac, and how a local app can read it]]></title><description><![CDATA[Disclosure: we are the team that makes OffiAI (Delicatino GmbH, Bremen). We usually have several agent sessions open at once: Claude Code in two terminals…]]></description><link>https://offiai.hashnode.dev/what-four-coding-agents-expose-on-your-mac-and-how-a-local-app-can-read-it</link><guid isPermaLink="true">https://offiai.hashnode.dev/what-four-coding-agents-expose-on-your-mac-and-how-a-local-app-can-read-it</guid><category><![CDATA[showdev]]></category><category><![CDATA[AI]]></category><category><![CDATA[macOS]]></category><category><![CDATA[devtools]]></category><category><![CDATA[privacy]]></category><dc:creator><![CDATA[OffiAI]]></dc:creator><pubDate>Mon, 16 Nov 2026 21:44:50 GMT</pubDate><content:encoded><![CDATA[<p>Disclosure: we are the team that makes OffiAI (Delicatino GmbH, Bremen).</p>
<p>We usually have several agent sessions open at once: Claude Code in two terminals, a Codex run, a Cowork conversation. Sooner or later one of them stops on a permission prompt in a window nobody is watching, and the work waits until someone notices.</p>
<p>To show them all in one place, we had to work out what each tool exposes locally. Below: four integrations, their limits, and every network connection the app itself makes.</p>
<h2>The common transport</h2>
<p>Every integration talks to the app on your own Mac: the Codex and Gemini hooks and the Cowork plugin over 127.0.0.1, with a port and a random token from a pairing file only your user can read; the app writes a new one each time it starts. The optional Claude Code approval hook uses a private socket file instead.</p>
<h2>Claude Code: transcripts, and permission prompts over stdio (no plugin, no hooks by default)</h2>
<p>To find sessions, the app runs <code>claude agents --json</code> and <code>claude auth status</code>, and reads the transcripts under <code>~/.claude/projects</code> (or under an absolute <code>CLAUDE_CONFIG_DIR</code>). It does not open Claude's credential or settings files. State follows the transcript: a prompt shows Thinking, a tool call shows the tool in use, the end of a turn shows Idle.</p>
<p>To start a task, the app launches <code>claude</code> in print mode in your project folder, with stream-JSON input and output and <code>--permission-prompt-tool stdio</code>, and writes your task description to its input.</p>
<p>That flag is the useful part. Every request that your Claude Code permission settings and the task's permission mode would ask about comes back over the pipe. The app shows the tool, its description and the command as written, with only credential-shaped strings masked. Allow runs the tool with its original input; Deny answers "Denied by the user in OffiAI."</p>
<p>A session started in a terminal creates no requests in the app unless you message it from the app or opt in. The opt-in is off by default; turning it on shows you the file path and the exact entry, then adds one <code>PermissionRequest</code> hook to Claude Code's user settings. Turning it off removes only that entry.</p>
<h2>Codex CLI: six hooks plus read-only rollout files</h2>
<p>An installer (<code>node install-codex.mjs</code>) adds six hooks (SessionStart, UserPromptSubmit, PreToolUse, PostToolUse, Stop, SessionEnd) to <code>hooks.json</code> in your Codex home, backs up the old file and keeps your other hooks; Codex asks you to approve them once. The hooks send lifecycle events, fixed labels, a prompt's word count and a hash of it, and tool names; never prompt text, tool arguments or output.</p>
<p>Status, model, token usage and Codex's latest answer come from the rollout files under <code>~/.codex/sessions</code>, read-only, with secrets masked. Codex's auth and config files are never opened.</p>
<p>To start Codex, the app runs it in a tmux session of its own (tmux ships inside the app), pastes your brief and reads the screen to tell when Codex is ready. "Open in Terminal" attaches you to the same session.</p>
<h2>Gemini (Antigravity CLI): a read-only database</h2>
<p>Nothing to install. The app opens <code>~/.gemini/antigravity-cli/conversation_summaries.db</code> read-only and lists the file names (not contents) in its <code>presence/</code> folder to see which conversations are running. It never runs <code>agy</code> to observe. Optional hooks add fixed lifecycle labels. Starting Gemini works like Codex, in its own tmux session, after asking you to trust the folder.</p>
<h2>Cowork: a plugin with a local MCP tool</h2>
<p>Cowork has no local transcripts to read, so here Claude does the reporting. The plugin (one file, or the <code>offiai/claude-plugins</code> marketplace on GitHub) adds a small local MCP server. Claude calls its <code>offiai_event</code> tool with a conversation label and a one-line summary; the bridge passes that, plus start, a heartbeat every 60 seconds and end, to the app. You see what Claude chooses to report, never a transcript.</p>
<h2>What we learned, and the limits</h2>
<ul>
<li>Not every integration can answer approvals. Allow and Deny work for Claude Code requests from tasks started in the app, terminal sessions you message from the app, and terminal sessions with the opt-in hook. Cowork requests are answered in the Claude app. Codex and Gemini ask in their own terminal, and a Codex session started there is read-only in the app.</li>
<li>A Codex session appears with your first prompt, because that is when Codex fires its hooks.</li>
<li>The Claude Code and Codex integrations support a specific range of CLI versions; sessions of other versions are not shown. The install pages list them: <a href="https://offiai.app/install/claude-code/">Claude Code</a>, <a href="https://offiai.app/install/codex/">Codex</a>.</li>
<li>Masking credentials, e-mail addresses and local paths before agent text is shown is best effort, not a guarantee.</li>
<li>The app runs no model of its own; each agent keeps its own account and provider.</li>
</ul>
<h2>Privacy model</h2>
<p>No account, no analytics, and no server of ours receives your sessions, prompts, answers or files, apart from the contents of a problem report you choose to send.</p>
<p>The network connections the app itself makes, and when:</p>
<ul>
<li><strong>Send feedback</strong>: only when you press Send, to offiai.app (retried later if you are offline). It holds your note, two screenshots (selected area and window), recent actions, the app log and state, and a random install ID created on your Mac. The conversation log is left out, but the screenshots show whatever is on screen.</li>
<li><strong>Installing or updating Claude Code</strong>: on your click, from Anthropic's download server.</li>
<li><strong>Installing Codex</strong>: on your click; npm fetches the Codex package, and Node.js comes from nodejs.org if no runtime is available.</li>
<li><strong>Signing in to Claude Code or Codex</strong>: on your click, the agent's own sign-in opens the provider's page in your browser.</li>
<li><strong>Robot ideas</strong>: only when you ask (or turn on Morning ideas), the app runs your own Claude Code once, without tools, in a private folder, with up to 20 masked session titles from the last 7 days and their state; it goes to Anthropic under your Claude account.</li>
<li><strong>Library questions</strong>: when you ask one, the app runs your own Claude Code in a private folder, with any sources you added.</li>
<li><strong>Network check</strong>: when you open it or press Check now, a DNS lookup and one HTTPS HEAD request each to the Anthropic, OpenAI and Google API hosts.</li>
<li><strong>Remote access</strong> from another device on your network is off until you turn it on. Opening another office, or an SSH tunnel you set up under Advanced, connects only to addresses you entered.</li>
</ul>
<p>The app does not check for its own updates. Your agents' traffic to their own providers is theirs. Full statement per integration: <a href="https://offiai.app/app-data/">https://offiai.app/app-data/</a>.</p>
<h2>Where this runs: OffiAI</h2>
<p>OffiAI is a free macOS app (Apple silicon only) that shows each session as a desk in a 3D office: give agents tasks at Reception, answer approval requests, review finished work. It is a public beta. The app is in English, Spanish and German (on a Mac set to another language it opens in English), and the <a href="https://offiai.app/changelog/">changelog</a> lists every build and what changed.</p>
<p>One-minute video: <a href="https://youtu.be/FiLY0iGKYUA">https://youtu.be/FiLY0iGKYUA</a> · Download: <a href="https://offiai.app">https://offiai.app</a></p>
<p>Our question: with several sessions running, which state do you most need to notice without looking: waiting for approval, finished, or stuck? Reply in the comments.</p>
<p>This article was written with AI assistance.</p>
]]></content:encoded></item></channel></rss>