Sidekick (macOS)

Theo Sidekick is a small Mac app with two workspaces. Transcription is system-wide dictation: hold a key in any text field, talk, let go, and Theo types the clean, finished text where your cursor is. Sidekick (beta) lets you ask Theo about anything on your screen — he answers out loud and points at the exact control. This article covers installing it, the permissions it asks for, and what each workspace does.

Install

Sidekick is a signed, notarized Mac app: one universal build that runs natively on Apple silicon and Intel (macOS 14 or later · Apple silicon and Intel), free with your OpenCharts account. Download it, drag it to Applications and open it. It keeps itself up to date from then on.

Download for MacmacOS

Everything it does is on the Sidekick page.

Permissions

On first launch, Sidekick walks you through a quick setup with one-click buttons. macOS requires you to grant each permission explicitly in System Settings → Privacy & Security.

Microphone

So it can hear you — dictation and hold-to-talk both need it.

Accessibility

So it can type where your cursor is, read the field it is typing into, and (in the Sidekick workspace) move the cursor and point at controls.

Screen Recording

Only for the Sidekick workspace: lets Theo capture a screenshot so he can answer about what is on your screen.

Speech Recognition

Lets you say “Hey Sidekick” to summon the assistant hands-free. Runs on-device.

Transcription only needs Microphone and Accessibility. You can re-check every permission any time from Settings → Access inside Sidekick.

The Sidekick workspace (beta)

Hold the orb and ask Theo about anything on your screen. It works, but it is still in beta — the pointing, clicking and typing can miss, and it improves in every release. Transcription is the part we consider finished, which is why a fresh install lands there.

Pixel-perfect pointing

Glides to the exact element, pulses a ring, and labels it. AX tree first, then vision fallback.

Voice push-to-talk

Hold the hotkey, ask out loud, hear the answer. Say “stop” to cancel, “repeat” to replay.

Step-by-step lessons

Teaches one step at a time. Say “next” to advance, with no extra wait between steps.

Per-app memory

Keeps conversation context separately for each application.

Inside the app, the What Theo can do chips on the Theo card explain each power and run a real sample question with Try it; the How it works pill in the title bar opens the full explainer with your own shortcuts filled in.

Talk and Text (Sidekick 1.78)

Since Sidekick 1.78 the Dashboard has a Text | Talk switch at the top, and every conversation is one or the other.

Two ways to have a conversation

Text is typed and read. Answers come back short and formatted: the answer first, then at most a few bullets or one small table. Theo stays quiet, and the mic beside the field writes your words into it so you can edit before you send.
Talk is a voice conversation: one big orb, one line that says what is happening, and the latest results under it. Tap to talk, tap again to end, or hold to ask one thing. The transcript is kept, one tap away under Show transcript.
The switch in a conversation’s header moves it between the two. Your keys, hold-to-talk and “Hey Sidekick” work as before.

Who answers

with Theo on the Dashboard, and the picker in a new conversation, list your personas under Theo. A persona answers as itself (its character, context and boundaries) with all of Theo’s powers, in Text and in Talk, in a Sidekick voice of its own that you can change in the picker. Build and edit personas in OpenCharts.

Sidekick 1.80 removed Theo’s browser, the signed-out cloud browser that 1.78 and 1.79 opened in a window beside the app, and the Where Theo works picker. Theo works on your Mac only: he sees your screen, works in your own apps and browser, answers current questions with a web search, and opens what he finds in your browser.

How a question travels

Every turn is the same four beats, whether you held a key, tapped one, said the wake word or typed. While a question runs, one quiet line under the Ask bar says which beat it is on, over a thin line that fills as the turn moves.

1
Listening
Your words are transcribed as you speak. A tap shorter than a third of a second, or a hold with nothing said, is ignored rather than sent.
2
Looking at your screen
One screenshot of the active window, taken the moment you finish (or of the screen or app you named, on any display). Theo reads it together with your question; sensitive windows are skipped.
3
Answering
The answer streams onto the answer card and is spoken sentence by sentence while the rest is still being written.
4
Pointing it out
When there is a control to show, the buddy flies to it and frames it with a ring. It goes home when Theo stops talking.

Where the answer lands, and how to stop it

The answer card sits at the bottom right of your screen: the transcript, the answer as it streams, tool progress and web sources. It fades a few seconds after a turn. Every question and answer is also kept in History.
Every conversation is a tab. Sidekick opens on Today. What you ask lands in a conversation tab beside it, with the whole thread; an ask from Today goes on with the conversation you were just having, or starts a new one after twenty quiet minutes. + (⌘T) starts a new conversation, History (⌘Y) lists every earlier one, on this Mac and in your OpenCharts account, grouped by day with a filter, and closing a tab (⌘W) never deletes it. ⌃Tab moves between tabs.
Just talk: pressing the orb or your key while Theo is speaking interrupts the answer and starts your new question.
Stop ends everything Sidekick is doing at once: the answer and its voice, the pointing, a task on your Mac (typing stops mid-sentence), a live conversation, a persona call and a Start my day guide. Say “stop”, or use Stop in the window’s header, on the answer card or in the menu bar. When Sidekick is busy while you are in another app, a small pill at the top of the screen says what it is doing and carries Stop. ⌘⇧Esc stops everything and sends the buddy home.
Turn off Sidekick (menu bar, the pill or the header) is Stop plus a switch on this Mac: no wake word, no shortcuts, no suggestions, no greetings and no calls until you turn it back on, even after a restart. Transcription keeps working.
Open your things by asking. “Open my Q3 launch board”, “play my notes movie” or “pull up the pricing deck” opens it in your browser right away. When several fit, Theo names them and you answer “the second one”. It works in a live conversation and inside a task too.
Pick up where you left off. When there is something to come back to, the big button on Today says what a press will do, with what is waiting under it. Finish it carries an unfinished task on at once from its last steps, even after a restart (“Stopped on a question · 12 min ago”): the press is your go-ahead. Reopen brings the tabs and documents back behind your browser, which comes forward once (“3 tabs · 1 document · 12 min ago”). Open brings back a conversation whose tab you closed. The card leaves once it is used. “Continue our last conversation”, “where were we” and the menu bar pick it up too.
Today’s plan. Once you have started your day, the day’s own button reads Today’s plan (beside the Continue card when both apply): your plan with what you already handled, kept up to date every minute, with anything new or moved marked under What changed. Say “today’s plan”, or use the menu bar or View → Your day (⌘3), to open it from anywhere.
Do it for me runs in the background. Theo says “On it.” and is free again at once, so you can keep asking or hand him another task. Up to four run at the same time, each with its own card in the conversation you asked in (what it is doing, its steps, how it ended) and its own Stop, and Today lists everything still going under Running now. Only one task uses the mouse and keyboard at a time, and on-screen steps wait while you are typing or moving the pointer. A step that deletes, sends or quits waits on its card for your go-ahead, and its conversation’s tab shows a question mark. A finished task puts one line in its conversation and says it once Theo is free; when one has to stop it says why: the connection, a rate limit, the account’s credits, a planner that took too long.
Tasks carry on. A step that cannot be done is skipped so the rest goes on, and the finished line says what was skipped. A task that stops on a question waits for your answer: on its card (with Continue and Skip that step), in your next words, or in a live conversation, and carries on from where it stopped. Asking for the same thing twice joins the running task instead of starting a second one.
Keep going in the background. While Theo works on an answer, press Keep going in background by the Ask bar, or end what you ask with “in the background” or “let me know when you’re done”. The answer carries on in its conversation while a fresh tab opens for what comes next, and Theo says “I have your answer on…” once he is free; an answer that keeps going does not look at your screen again. Documents Theo makes and questions to the sites you connected always work this way: Theo says he is on it, the chip lands in their conversation when it is ready, and a document opens in your browser only while you are looking at that conversation.
Close exactly what you asked. “Close my Research tab group” closes that group, “close the Gmail tab” that one tab, and a window closes by its own close button. “Close the Chrome windows” closes them all in one step, each through its own close button and each checked, naming any that asked to save. Your own words decide: closing never quits the app unless you say quit (even in a live conversation), a quit is checked against the running apps rather than assumed, and shortcuts that quit or close a whole window always ask first.
Every screen. A task sees the windows on all your displays (up to eight), brings the one it needs forward wherever it is, and remembers what it opened, so “close what you opened on the other screen” works. Ask “what’s on screen 3?” or “read me what Slack says” and Theo looks there.
“Remember that…” saves the fact at once and adds a dated line to a pinned Scratchpad note called Theo remembers, then Theo confirms. The same works in a live conversation.

Ways to talk to Theo

Five ways in, one pipeline. The keys are yours to change under Settings → Sidekick; the app keeps the ask-Theo keys separate from the dictation keys so the two never fire together.

Hold the orb

Press and hold the orb on the Theo card, ask, let go. One question, one answer.

Hold your hold-to-talk key

The same thing from any app. The first syllable is kept even if you start talking as you press.

Tap your hands-free key

Opens a live conversation when it is on in Settings, otherwise a single question that sends itself after you pause.

Say “Hey Sidekick”

Recognized on your Mac by the system’s own speech recognition; that audio never leaves the machine. Off by default; switch it on in Settings.

Type

Ask in the field on Today or in any conversation tab. Typed questions still look at your screen.

One question or a conversation? A hold is one screenshot, one answer, done: best for a quick fact or a where-is-it. A live conversation keeps listening, interrupts itself when you speak, can glance at your screen every few seconds, and ends when you say goodbye or tap End. A pause is part of a conversation: after a minute of quiet Theo asks once whether you are still there, and after two it ends. A live conversation has a ten-minute limit, but it no longer ends there: shortly before it Theo writes himself a summary of what was said, and at the limit he carries on in a fresh session with that summary in mind, no greeting, no recap, the microphone still on, up to about an hour in all. Only a conversation that has gone quiet ends at the limit, with a heads-up a minute before.

Personas in Sidekick

Talk to a persona you built, on your desk

A Theo Persona is a live video agent you build in OpenCharts: a face, a voice, a character, its knowledge, its boundaries and the tools it may use. The Personas card on Sidekick’s Home lists yours; Talk opens a small always-on-top call window next to whatever you are working on, with the persona’s video, live captions and a chip for each tool it runs.

Everything the persona can do, it can do here: the same knowledge, the same tools, and Agent Link when it is switched on, so your own coding agent answers through it. The card says honestly whether that agent is connected right now.
Your recent personas first. The card and the Start my day picker show the three personas you called most recently; the rest are under More. Calls from visitors on a published link never change the order.
One conversation at a time. Starting a call interrupts anything Theo was saying, ends a live conversation, and pauses the wake word for the whole call; starting another call hangs the first one up.
Your microphone joins, your camera stays off until you turn it on. Mute, camera, a shrink-to-bubble button and End are in the window; Stop anywhere in Sidekick also ends the call.
The flask button places a test call that bills no minutes; a real call is billed like one started from the Studio and appears in the persona’s activity with its channel.
Personas need a Pro or Teams plan and a persona that is published with a face; the card explains what is missing and opens OpenCharts to fix it.
Your Mac. While you are on a call from Sidekick, the persona can look at your screen, point at a control, click one, or carry out a task through the app, then tell you what happened. Calling from the Personas card switches this on for that persona by itself (the tile’s Your Mac switch shows it), and the call runs as your own private call even on a persona you published publicly. Steps that delete, send, pay or quit wait until you say yes; password, banking and sign-in windows are never looked at; and when you are not on a call from Sidekick the persona says it cannot reach your Mac.
It knows where you are. As you move between apps during the call, the persona gets a short note about which app and window you are in, never read aloud, so “what is this page about?” means something. Sidekick comes onto your screen as the call connects, ready to point.
The call window is the persona’s video edge to edge: move the pointer and the name, Fullscreen, Shrink and End fade in at the top; what is said appears on a soft gradient at the foot. Esc leaves full screen; Shrink folds the call into a small bubble. Ending the call, however you end it, closes the window.
Dictate during a call. Hold your dictate key and the call hands the microphone to Sidekick: the persona pauses and does not hear what you dictate; when you let go it hears you again.
Results, right there. As the persona works, cards slide into the foot of the window: the tool that is running, a web answer with its sources, the places it found, a map card with Open in Maps, your agent’s reply, and anything it builds with Open in OpenCharts the moment it exists.
Build from your screen. Say “make a deck out of this page” or “turn this into notes” and the persona reads what is in front of you first, then builds from it; the result lands as a card in the window.
The Mac’s answer always reaches the persona. Ask it to open a site and it goes there first; a task that takes a while is handed to the persona the moment it finishes, and while it runs the persona can still look at your screen for you (a second task hears “still on the last one”).

In Sidekick 1.78 personas answer right in your conversations, on Theo’s engine (see Who answers); the Personas card and its test calls are gone, and the call window is used only when a persona guides your Start my day.

Setting Sidekick up. The first run asks near its end whether you would like to set up Sidekick; say Later and the Sidekick tab shows a setup screen with the permissions it needs (Screen Recording, Speech Recognition, plus the Microphone and Accessibility Transcription already uses) until you press Start using Sidekick. With no personas yet, the Personas card explains what one is and opens OpenCharts to create the first.

Transcription: talk instead of type

A second workspace with its own shortcuts

Switch the window to Transcription and Sidekick becomes a dictation tool for every app on your Mac. Put your cursor in any text field, hold the fn key (or your own combo), say what you mean, and let go: Theo removes the fillers, fixes grammar and punctuation, matches the tone of the app you are in, and pastes the finished text where your cursor is. One ⌘Z undoes it.

Hold to dictate, double-tap to keep recording hands-free, Esc to cancel. Up to four shortcuts per action, kept separate from the ask-Theo keys.
Cleanup levels from None to High, and a tone per destination (personal messages, work chat, email, notes, documents) — Auto picks one for the place you are in, or choose Formal, Casual, Very casual or Excited yourself. An email dictated into a document is written as an email; a subject line gets one line.
Dictionary for names and jargon (with “heard as” fixes and auto-learning from your corrections — or click a word in History, type what you meant, and Theo learns it on the spot), Shorthand that expands a spoken cue into saved text, and a Scratchpad for thinking out loud.
Edit by voice. Select text, hold your dictate key and say the change (“make it shorter”, “change Tuesday to Monday”, “spell it T-O-N-I”) and Theo rewrites the selection in place; dictating plain words over a selection still replaces it. For a minute after a paste, with nothing selected, “make that shorter”, “change five to six” or “undo that” fixes what just landed, right where it is. Both switches live under Transcription → Settings → Edit by voice.
Command Mode and Transforms: select text, say what to do with it on its own key (or press a saved transform’s shortcut) and Theo rewrites it in place.
Send with your voice. End a dictation with “send it” (or “press enter”) and Theo presses the send key after pasting; said on its own, it sends what is already in the box. Only as the last thing you say: “I’ll send it tomorrow” stays text. Off by default, with the key per destination (Return in a chat, ⌘Return in an email) under Settings → Experimental.
Re-run as… on any History row: the same words again through Polish, Email Master, Writer or one of your translate pairs. The result goes to your clipboard and the paste-last key, never straight into an app.
Hear the names on screen. With Context awareness on, the people and products visible around your cursor, in the window title and, in a code editor, the names in the code are handed to the speech engine for that one dictation, so they are heard spelled right without a dictionary entry. In Slack, Messages and Mail the sender’s and recipients’ names count too. Never from a password or banking window.
Team dictionary. On the Teams plan, owners and admins keep one shared list of names and products, and everyone on the team dictates with it, marked with the team glyph in Dictionary. Up to 500 words, CSV import included.
On this Mac. The on-device engine transcribes without the cloud, and on macOS 26 with Apple Intelligence, Clean up on this Mac too runs the Light and Medium cleanup on Apple’s own model, so a dictation can land without a word leaving the machine. High cleanup, a take past 400 words, Command Mode and transforms still go to Theo. Turn other audio down while I talk lowers the Mac’s volume to a murmur while you hold the key and puts it back the moment you let go.
The Voice Bar shows what Theo is hearing — a minimal orb that rises from the bottom edge of the screen (a sliver at rest, a dome with a waveform while you talk), or the full status bar with a timer and buttons; your pick in Settings. Transcripts stay on your Mac; dictionary, shorthand and styles sync.

Dictation needs the Accessibility and Microphone permissions, and works best when macOS’s own “Press 🌐 key to” setting is set to Do Nothing.

Command line and MCP

Your dictations, from the Terminal and from your agents

Sidekick ships a small command-line tool that reads what the app already keeps on your Mac: your dictation history, dictionary and shorthand. It is read-only: nothing it does can change a dictation, send one anywhere or reach the network. Set it up once from Transcription → Settings → Command line & MCP, which shows the command that puts sidekick in your Terminal and the one that adds it to Claude Code.

sidekick search budget review finds dictations by their words; sidekick recent --since 2026-10-01 lists the latest; add --json to either for machines.
sidekick export --format md (or txt, jsonl) writes your history out; sidekick stats --days 7 is Insights in one screen: dictations, words, speed, streak, where the words went.
sidekick dictionary and sidekick shorthand print your words (and the team’s) and your saved cues; sidekick paths says where the files are.
sidekick mcp runs an MCP server over stdio with the same five reads as tools (search_dictations, recent_dictations, dictation_stats, dictionary, shorthand). Add it to Claude Code with the command in Settings, or point any other MCP client at the same program with the one argument mcp; the agent can then look through your dictations when you ask it to, on this Mac only.

Adding words, shorthand or team entries stays in the app: a command line that could edit the dictionary would need your account’s credential, and that belongs in one place.

Sign in with OpenCharts

No API keys to paste

Sidekick is powered by Theo, through OpenCharts. Click Sign in with OpenCharts and it opens your browser to log in, then connects on its own and stays signed in across relaunches and updates. Each Mac you sign in on appears as its own computer under Settings → Security → Sessions on the web, where you can unlink it. After that, dictation, vision pointing and spoken answers all run through Theo.

Auto-update

Always on the latest version

Sidekick checks for updates on launch and verifies each one against an embedded public key, so only releases from our build pipeline can install. You can also check any time from Settings → Account → App updates, which shows your current version and an Install & Restart button when a new build is available.

Privacy

A built-in privacy guard

Sidekick skips the screenshot when a password manager, banking, or login window is in front.
It also skips the screen for “who is…” identity questions.
Screenshots are sent for a single request to answer your question and are never logged or persisted.

Was this article helpful?

Theo Desk (macOS)

Native macOS desktop agent that lets Theo see your screen, drive any application, and run multi-step workflows on your Mac.

Theo AI Overview

Overview of Theo AI — 11 chat modes, creation tools, generative UI, memory, and the sidebar mascot.

Voice Mode

Real-time streaming voice conversations powered by Theo Voice — talk to create projects, research, and more.

Learn Mode

Orchestrated learning suite — flashcards, quizzes, study notes, and presentations generated in parallel from any topic.

Theo Desk (macOS)Voice Mode