How to use Ask Agent with your coding agent
Hold the Ask Agent key, ask out loud, and the coding agent you already use — Claude Code, Codex, Gemini CLI, OpenCode, Hermes, OpenClaw, Grok, or any command of your own — answers from your own computer. By default it works quietly with no window and Walkie reads the reply back in its Read Aloud voice; flip one switch and it opens in its own terminal window instead so you can watch. Set it up in Settings → Agent: switch on Enable Agent, pick who answers under Answers come from, and check the line underneath says where Walkie found it. With Claude Code as your agent, it can also hand work to a coding session you already have open, and answer calendar questions in a few seconds. Ask Agent needs Walkie Pro or above and Walkie 1.7.3 or later.
Ask Agent is Walkie's ask-a-question key. You hold it, say something, and an agent answers — Walkie shows its speaking waveform and reads the answer aloud so you never leave what you were doing. In Cloud Mode, out of the box, the answer comes from Walkie's own assistant. In On-Device Mode that hosted assistant is refused — point it at Claude Code, Codex, or another agent on this computer, or it will not send the question to Walkie. On Omarchy it follows the agent the machine designates. This page is about pointing it at a coding agent: the one on your computer that can actually open files, run commands and change settings. You say "what's eating my disk space?", the agent looks, and Walkie tells you.
Which agents work
Everything in the Answers come from list, in the order it appears. Agents Walkie cannot find on this computer are still listed, greyed out with "(not installed)" next to them.
| Option | What it is | Can it change things on this computer? |
|---|---|---|
| This computer's agent (Omarchy only) | Whichever agent you designated with omarchy default agent. Walkie follows that choice, so changing it in Omarchy changes it here too. | Yes |
| Claude Code | Anthropic's command-line agent. | Yes |
| Codex | OpenAI's command-line agent. | Yes |
| Gemini CLI | Google's command-line agent. | Yes |
| OpenCode | The open-source command-line agent. | Yes |
| Hermes | Nous Research's agent. | Yes |
| OpenClaw | The OpenClaw agent. | Yes |
| Grok | xAI's command-line agent (Grok Build). | Yes |
| Walkie's assistant | Walkie's own cloud assistant, with Walkie's connectors and tools (weather, news, web search, your connected apps). | No commands, no files. It can open an app or a link you have allowed, paste text and draft an email, and that is all |
| Custom command… | Any program you name that takes the words and prints an answer. See Using your own command. | Whatever that program can do |
Every coding agent runs with its own account, its own tools and its own permissions — exactly as it would if you typed to it in a terminal. Walkie's connectors are not involved while an outside agent is answering; the Connectors section in Settings → Agent says so while one is selected.
Set it up
- Install the agent the way its maker documents, and make sure it works in a terminal on its own first.Walkie only hands over the words. If the agent is not signed in or is not installed, there is nothing for Walkie to hand them to.
- Open Walkie and go to Settings → Agent. Switch on Enable Agent.If the page shows an Experimental notice instead, click the button on it first — the Agent lives behind that toggle.
- Under Answers come from, choose your agent.Installed agents are listed by name. One marked "(not installed)" is one Walkie could not find — see Why is my agent greyed out? below.
- Read the status line under the list.It says where Walkie found the agent — for example "Found at ~/.nvm/versions/node/v22.16.0/bin/claude" — or that it cannot find it. If it cannot, Walkie's assistant answers until that is fixed.
- Check the Ask Agent shortcut on the same page, then hold it and ask something.The defaults are below. Click the shortcut field and press a new combination to change it.
The Ask Agent key
Any launcher, script or window manager can trigger Ask Agent too, on every platform: run walkie --quick-ask. One press is enough — Walkie notices when you stop speaking.
Hear it, or watch it
There are two ways a coding agent can answer, and a switch on the settings page picks between them. It appears whenever an installed outside agent is selected — on Omarchy, for This computer's agent only when that agent is one Walkie can run quietly.
| Read back (default) | Open the agent's window instead | |
|---|---|---|
| What happens | The agent works in the background with no window. Walkie waits for it to finish and reads the final answer aloud. | The agent opens in its own terminal window with your words already typed, and you watch it work. Walkie flashes "Sent to …" and stays quiet. |
| Good for | Quick questions and small jobs where you want the result, not the process. | Longer jobs, anything you want to steer as it goes, or when you would rather see the agent's own screen. |
| Follow-up questions | Ask again in the same Ask Agent session and the agent continues its most recent conversation, so "and delete the biggest one" makes sense after "what's eating my disk space?" | Keep typing or talking in the window that opened. |
Where the window opens: on macOS in Terminal (or whichever app you have set to open .command files, such as iTerm2 or Ghostty); on Windows in the agent's own console window; on Linux through Omarchy's launcher when it is present, otherwise your desktop's default terminal. On Omarchy with This computer's agent selected, the window path is Omarchy's own omarchy-agent-prompt launcher.
How it talks back
Ask Agent answers the way a person would on a call, not the way a terminal prints. A question is answered within a couple of seconds and spoken sentence by sentence as the reply is written. A request that needs the computer gets a spoken "On it" at once, and the agent takes over; while it works, the panel above the pill lists what it is doing and thinking, and the words it says are read aloud with the highlight as they arrive. Replies are shaped for speech — short, plain, no headings or bullet points read out — and the full text stays in the panel and in your history.
With Claude Code chosen, a quick question goes to its fastest model with no tools, on your own plan; only a request that needs the computer runs the full agent. Press ✕ on the panel or Esc to stop the agent and the voice together.
When the agent ends on a choice — "Which of these do you want?" with a list, or a yes-or-no question — the panel shows the choices as buttons. Click one, or just say it ("the second one", "yes"), and that becomes your next turn.
Say "go back to my last conversation" and the panel lists your recent conversations to pick from, by voice or by click. With Claude Code or Codex, Walkie brings back the agent's own session too, so it remembers what it did, not just what was said.
The panel's Open in Claude Code (or Open in Codex) button opens the whole conversation in a terminal, resumed on the agent's own session, when you want to see everything the panel only summarised.
Send work to a session you already have open
With Claude Code as your agent, Walkie works out what you are asking before it does anything with it.
| What you ask | What happens |
|---|---|
| A question | It simply gets answered. |
| Something that needs your accounts or live information, like an email | A fresh agent takes it on. |
| Work on one of your projects | Walkie asks "Send this to a session already open?" and lists the open sessions it could belong in. |
Pick the session — click it or say it — and Walkie brings that window to the front, types your request in and presses Return, then tells you it did. Typing it in works in Terminal on a Mac.
Quick calendar answers
Ask "what's on my calendar tomorrow?" and Walkie answers in a few seconds from the Google Calendar you connected for Meetings, instead of waiting on your agent. Turn on Quick calendar answers in Settings → Agent. It is off by default, and it works when your agent is Claude Code.
Things to ask
Anything a coding agent could do if you typed it. Short and concrete works best — say what you want, not how to do it.
- "What's eating my disk space?"
- "Make the terminal font bigger."
- "Which of my Homebrew packages are out of date?"
- "Summarise the git log for this week in my Work folder."
- "Turn on dark mode at sunset."
- "Rename every screenshot on my desktop with the date it was taken."
- "Add a battery widget to the bar." (Omarchy)
- "Why is my fan running? Find what's using the CPU."
Using your own command
Custom command… runs any program that takes the words and prints the answer — a script of your own, an agent Walkie does not list, or a listed agent installed somewhere unusual. Type the command line in the field that appears:
hermes chat --yolo --query={prompt}- Put
{prompt}where your words should go. Leave it out and they are added as the last argument. - It runs from your home folder (or your Work folder) and never through a shell — so your words can contain anything. No window, unless you turn on Open the agent's window instead, which works for a custom command too.
- If the program is not something Walkie can find by name, give its full path:
/Users/you/.local/bin/my-agent. A full path always works.
What it costs, and where your words go
Two different services are involved, and each bills its own part.
| Who pays | Where the data goes | |
|---|---|---|
| The agent's thinking | The agent's own plan or API key — your Claude, ChatGPT or Gemini subscription, or whatever you set the agent up with. Walkie does not bill for this. | To that agent, and wherever it sends things. Its privacy policy applies, not Walkie's. |
| Walkie's part — hearing your question and speaking the answer | Counts against your Walkie plan exactly as any other dictation and Read Aloud does. See Plans and limits. | In Cloud Mode, your question is transcribed and the reply is voiced through Walkie's cloud. In On-Device Mode both happen on this computer, and nothing leaves it except what the agent itself sends. |
| Walkie's assistant (when that is what answers) | Your Walkie plan. | Walkie's cloud. See What leaves your device. |
What Walkie keeps the agent away from
When Walkie starts Claude Code, it starts it locked down. A request can carry text Walkie did not write — a calendar invite, a web page, a file the agent opened — and that text could try to talk the agent into something you never asked for.
- Saved passwords and keys. The agent is blocked from reading the usual places saved passwords, keys and browser logins are kept, and Walkie's own sign-in.
- Links and web pages. Before it fetches a web page or opens a link, Walkie shows an Allow button, and nothing opens until you press it. When the agent is working in its own window, it asks there instead. Searching the web does not ask.
On Omarchy
Omarchy designates one agent for the whole machine with omarchy default agent claude (or codex, opencode, hermes, openclaw, grok, and others such as Pi, Crush and Cursor CLI). Walkie reads that choice and offers it as This computer's agent. When the designated agent is one Walkie can run quietly — Claude Code, Codex, OpenCode, Hermes, OpenClaw or Grok — you get the read-back by default and the window toggle to change it; otherwise Walkie opens Omarchy's own agent window with your words. The keys and the bindings.lua block are in Walkie on Omarchy.
Frequently asked questions about how to use Ask Agent with your coding agent
Why is Claude Code (or another agent) greyed out with "(not installed)"?
Walkie could not find its program. It looks in the places Walkie itself can see, then (on macOS and Linux) asks your login shell for its search path once at startup — which is how installs from nvm, fnm, mise and Volta are found even though apps launched from the Dock see a much shorter path — and then checks the usual install folders (Homebrew, npm and bun globals, ~/.local/bin, Claude Code's own installer, and their Windows equivalents). Three things to try, in order: confirm the agent runs in a terminal at all; click back into Walkie's window, because it looks again every time the window regains focus — and if the install added a new folder to your shell's search path, quit and reopen Walkie, since that path is read once at startup; and if it is still greyed out, choose Custom command… and type the program's full path, which skips the search entirely.
Does it need Walkie's connectors?
No. The agent brings its own tools, MCP servers and permissions — whatever you set up when you configured it. Walkie's connectors only take part when Walkie's assistant is the one answering, and the Connectors section in Settings → Agent says as much while an outside agent is selected.
Can it change things on my computer?
Yes. That is the point, and it is also the risk. Walkie launches each agent in its don't-stop-to-ask mode, so a request like "clean up my downloads folder" is carried out, not proposed. The agent uses the account and permissions you gave it. If you want answers without any of that, pick Walkie's assistant, which runs no commands and changes no files — it can only open an app or link you have allowed, paste text and draft an email. With Claude Code, Walkie also keeps the agent away from the usual places passwords and keys are kept, and asks you to Allow it before it opens a link — see What Walkie keeps the agent away from.
Why did Walkie copy my request instead of typing it into my session?
Because it couldn't be sure it had the right window. Walkie types into a session only when it can find that window and bring it to the front, which works in Terminal on a Mac. When the session is inside a code editor, or the window won't come forward, the request goes on your clipboard and Walkie tells you, so you can paste it in yourself.
Does a question about my calendar go to my agent?
Not when Quick calendar answers is on in Settings → Agent and Claude Code is your agent. Then Walkie reads the Google Calendar you connected for Meetings and answers in a few seconds, sharing your upcoming events with Claude to answer it. With the setting off, a calendar question goes to a fresh agent like any other request that needs your accounts.
How do I trigger it from a script or a window manager?
Run walkie --quick-ask (on macOS the program lives inside the app: /Applications/Walkie.app/Contents/MacOS/walkie --quick-ask). It tells the running Walkie to start a Ask Agent and returns straight away; Walkie stops listening on its own when you finish speaking, so one binding is enough. On Wayland compositors, where an app cannot register its own global hotkey, this is how the key works at all.
What if the agent takes a long time?
In the read-back mode Walkie waits up to 10 minutes. If the agent has not finished by then, or exits with an error, Walkie shows its error overlay briefly and nothing is spoken — Walkie's own log records the reason, including the last lines the agent printed. For jobs you expect to run long, turn on Open the agent's window instead and watch it work with no time limit.
Does it work on Omarchy?
Yes, and Omarchy is where it started. Pick the machine's agent with omarchy default agent, run Walkie's Set up my keys, and F8 or Super + Alt + M starts a Ask Agent. Walkie runs the designated agent quietly when it can and opens Omarchy's agent window otherwise.
Can I answer the agent without speaking?
Yes. When the agent asks a question that has a list of options or a yes-or-no answer, the panel shows the options as buttons. Clicking one sends it as your reply, the same as saying it.
Can I pick up an earlier conversation?
Say "go back to my last conversation" (or "continue the previous chat") and choose from the recent conversations the panel lists. With Claude Code or Codex the agent's own session comes back with it.
How do I end a conversation?
Say so. When Walkie asks whether there is anything else and you answer "No, I'm good", "that's all", "thanks" or "bye", it signs off, the panel comes down and it stops listening. Esc or the ✕ on the panel does the same at any point, and also stops an agent that is still working.
Why does macOS ask whether Walkie can access my Downloads, Documents or Music folder?
Because the agent runs inside Walkie. When you ask it to do something that looks in those folders — "what is eating my disk space", "rename the screenshots on my desktop" — macOS protects them and asks on Walkie's behalf, the same way it would for Terminal. Walkie itself never opens those folders. To stop the per-folder questions altogether, give Walkie Full Disk Access once: Settings → Agent → Files and folders shows whether it is granted and opens the right pane in System Settings → Privacy & Security, and the agent panel offers the same the first time a task needs your files.
Which voice reads the answer?
The same voice Read Aloud uses. In Cloud Mode that is the Voice picked in Settings → Agent, one of Walkie's cloud voices; in On-Device Mode it is the on-device voice chosen in Read Aloud's settings, and the answer never leaves this computer. See Reading text aloud.