Your coding agents,
out loud.
Pintail reads Claude Code’s replies out loud as a short summary, with live captions in the notch, so you can leave the desk and still know which session finished.
$15 once · 20 free read-outs · no subscription
> fix the empty state on the review screen
Update(components/EmptyState.tsx)
Bash(pnpm test) ✓ 14 passed
Fixed. The empty state never showed when a filter matched nothing: Review.tsx checked items.length before applying the filters, so a filtered-out list rendered as a blank card with no explanation. While tracing it I found two related problems, so here is the full picture.
Root cause:
- useReviewItems() returns the unfiltered list, and the filtering happens later in ReviewList via applyFilters(items, query). The empty check in Review.tsx ran on the unfiltered list, which is almost never empty, so the EmptyState branch was effectively dead code once a user had any items at all.
- isLoading starts as false on the first render because the query is created lazily in a useEffect. For one frame the list is empty and not loading, so the empty state flashed before the skeleton appeared. That is the flicker reported in #412.
- The empty-state copy lived inline in Review.tsx, and a slightly different version lived in Archive.tsx, so the two screens had drifted apart ("Nothing here yet" vs "No reviews to show").
What I changed:
- screens/Review.tsx: moved the filtering into a useMemo next to the query, and the empty check now runs on the filtered list. I also create the query eagerly, so isLoading is true on the first render and the flash is gone.
- components/EmptyState.tsx: the component owns the copy now. It takes a variant prop, "no-items" for a new inbox and "no-matches" when the filters hide everything, with a Clear filters button in the second case that resets the query params through useFilters().reset().
- hooks/useFilters.ts: added reset(), which clears all URL params in one navigation instead of one per filter, so the back button no longer walks through every intermediate filter state.
- screens/Review.test.tsx: three new tests. An empty inbox shows the no-items variant; a filter with no matches shows the no-matches variant and Clear filters brings the list back; the first render shows the skeleton, not the empty state.
Verification: pnpm test passes, 14 tests in 3 files. pnpm typecheck and pnpm lint are clean. I did not run the Playwright suite because it needs the staging API, but none of the e2e specs touch the review screen.
Not changed: Archive.tsx has the same pattern and the same dead empty branch; I left it alone to keep this change focused. The filter chips also do not announce their state to screen readers, which is unrelated but worth a follow-up.
Want me to apply the same fix to the Archive screen next, or open an issue for the accessibility gap first?
The empty state is fixed and all 14 tests pass.
Six tabs. A little less chaos.
I run five or six agent sessions at once, and they finish whenever they like. When a reply lands, Pintail says the short version and which project it came from, one at a time. A pill at the top of your screen shows the words as they’re said, and a little pet on your desktop shows who’s talking. Reading them automatically is optional.
Try it: switch tabs, then click the pet to hear the tab you’re in.
Illustration: a MacBook with iTerm2 and six Claude Code sessions. Three finish while you work in review-app, and Pintail reads each reply out in the order it came in. Try it: switch tabs, then use the pet button to hear the tab you’re in.
> the review screen is blank when a filter returns nothing, why?
Read(screens/Review.tsx)
Read 186 lines
Read(hooks/useReviews.ts)
Read 44 lines
The list renders nothing when items is empty, and there’s no fallback. useReviews also returns an empty array while it’s loading, so the screen can’t tell “loading” from “no results”.
I’d have useReviews return a status (loading, empty, ready) and show a proper empty state for empty. Want me to do that?
> yes
Update(hooks/useReviews.ts)
Updated hooks/useReviews.ts with 9 additions and 4 removals
Update(screens/Review.tsx)
Updated screens/Review.tsx with 21 additions
Bash(pnpm test)
Tests: 198 passed, 198 total
Done. useReviews returns a status now, and the review screen shows “No reviews match this filter” with a Clear filters button when it’s empty. Loading still shows the skeleton. Tests pass.
> add empty states to every list screen, and make them match the review one
Search(pattern: "<List", path: "screens")
Found 7 files
Write(components/EmptyState.tsx)
Wrote 48 lines to components/EmptyState.tsx
Update(screens/Inbox.tsx)
Updated screens/Inbox.tsx with 12 additions and 3 removals
Update(screens/Projects.tsx)
Updated screens/Projects.tsx with 10 additions
Bash(pnpm test)
Tests: 212 passed, 212 total
Every list screen now has an empty state, and they all use the same component. Here’s what changed, screen by screen:
- Review: I moved its empty state into components/EmptyState.tsx and made the title, text and button props.
- Inbox: “Nothing new” with a button to the archive, and its own message when a search finds nothing.
- Archive: “Nothing archived yet”. No button, there’s nothing useful to do from there.
- Projects: a Create project button, the same one as in the header, so there’s one clear next step.
- Team: explains that invites go out by email and links to the invite form.
- Tags: “No tags yet” and one line on what tags are for.
- Search: repeats the query and offers Clear search.
Two screens showed a header with nothing under it while loading, which looked broken. They now show the loading skeleton until the first fetch finishes.
I added a test for each empty state and one for the search case, 14 in total. The whole suite passes.
Two things to check: the wording is a first draft, and the Team link goes to /settings/invites, which only admins can open.
>
? for shortcuts
Less reading. Same conversation.
Agents write a lot. Pintail reads the reply and says the short version. Pick how much you hear, from one sentence to the whole reply. The full text stays in your terminal either way.
> why do people get logged out after a deploy?
I looked at the auth flow end to end. Here is what happens. The session cookie is signed with SESSION_SECRET, and the deploy script generates a new secret on every release. Every cookie signed with the old secret fails verification, so the middleware treats the user as logged out and clears it. It only shows up after deploys because that is the only time the secret changes. Local dev never rotates it. I moved the secret into the environment config so it stays the same across releases, added a fallback that accepts the previous secret for 24 hours, and wrote two tests: one for a cookie signed with the old secret, one for the new one. Both pass, along with the existing 212 tests. You may want to set SESSION_SECRET in production before the next deploy.
Fixed, just set SESSION_SECRET before the next deploy.
Deploys changed the session secret. Fixed. Set SESSION_SECRET before the next deploy.
Every deploy made a new session secret, so people got logged out. Local dev never rotates it, which is why you only saw it after deploys. The secret now stays the same across releases, and the old one still works for a day. Two new tests cover it, and all 214 pass. Set SESSION_SECRET in production before the next deploy.
I looked at the auth flow end to end. Here is what happens. The session cookie is signed with SESSION_SECRET, and the …
Set it up once. Then listen.
1 Connect your coding agent.
Pintail picks up replies from your terminal sessions through hooks. Codex asks you to trust them once.
2 Choose your voice setup.
Paste your own OpenRouter key, or buy Pintail credits in the app and skip the account.
3 Ask, then listen.
Hear a short summary with the project’s name and live captions. Automatic reading is optional.
Pick your companion.
Four pets, one job: showing who’s talking. Yes, one is a rubber duck. You spent years explaining your code to it; now it tells you what your agent did. Click one to hear it. The rest of the page follows your pick.
$15 once
- 20 free read-outs to try it first
- No subscription. No account.
- For Macs with Apple silicon
Not included: OpenRouter usage. The voice runs on your own OpenRouter account, usually ½–2 ¢ a reply.
Fair questions.
How do I make Claude Code read its replies out loud?
Pintail reads Claude Code’s replies out loud as a short summary, with live captions in the notch, so you can leave the desk and still know which session finished. It costs $15 once, includes 20 free read-outs, and needs no account. Automatic reading is off until you turn it on.
Is there text to speech for Claude Code on a Mac?
Yes. A free Stop hook can pipe the reply into the Mac’s system voice. Aloud reads it on the Mac with Kokoro. Pintail speaks a short summary and puts the project on the notch. The alternatives page compares them.
Does Claude Code’s voice mode read replies out loud?
No. Claude Code’s /voice is dictation: you speak, it types your prompt. Anthropic’s help page says that in Claude Code, dictation is available but voice mode isn’t. Pintail is the other half: it speaks the reply back.
How do I get notified when Claude Code finishes?
Claude Code shows a desktop notification in Ghostty, Kitty and iTerm2, and a Stop hook can play a sound in any terminal. Neither says which session finished or what it did. When a reply finishes, Pintail speaks a short summary and the notch pill names the project, with live captions, so you know even if you left the desk.
What tracks several Claude Code sessions at once?
Pintail uses one voice for every tab. Replies play in the order they finished, and the notch pill shows which project is talking. Heard, Agents Aloud, SpeakySpeak and Kobi also keep sessions apart; the alternatives page says how.
Which coding agents does it work with?
Claude Code and Codex, in the terminal and in their desktop apps.
There are free tools. Why pay for this one?
If you only want a free voice, try Aloud or SpeakySpeak; I compare them on the alternatives page. Pintail is for three things they don’t do together: a short summary instead of the whole reply, several agents at once without them talking over each other, and always seeing who’s talking, with the pill’s live captions and the duck. It has a free local voice too.
Is it a dictation app?
No, it’s the other half. You dictate with whatever you like (I use FluidVoice, and Claude Code has /voice). Pintail speaks the answer back.
Does it use my Claude tokens?
No. The summary is written outside the agent, and the voice commands you type in the prompt (like “voice read”) are caught by the hook and never reach the agent.
Do I need an OpenRouter account?
No. Paste your own OpenRouter key and pay OpenRouter for what you use, or buy Pintail credits in the app and skip the account.
What does the voice cost?
I measured it in September 2026: about 1.5 to 3 cents per 1,000 spoken characters, depending on the voice. A spoken summary is a few hundred characters, so usually half a cent to 2 cents a reply. OpenRouter also has free voices with daily limits. Check OpenRouter for current prices.
Can it run without the cloud?
Yes, in English. The voice can run on your Mac (Kokoro, a free 380 MB download), and on macOS 26 Apple Intelligence writes the summary there too. Then nothing leaves your Mac. Other languages use the cloud voice.
What gets sent where?
To speak, Pintail sends the text it reads to OpenRouter, through your account. The short version can be written on your Mac with Apple Intelligence. If you pick a cloud summarizer instead, the reply goes to OpenRouter too. Your replies never go to us.
Does it read every reply out loud?
Only if you want it to. Automatic reading is off by default; you can turn it on, or play a reply when you’re ready.
Which languages?
English and German. Spanish, French, Italian, Dutch and Portuguese are in beta; I’ve listened to English and German.
What will it cost?
$15 once, with 20 free read-outs to try it first. No subscription. OpenRouter usage is separate.
Which Macs?
Macs with Apple silicon.
Why I built it.
I’m Kevin. I tend to have five or six Claude Code tabs going at once. I already dictate my prompts with FluidVoice. Then the agent replies with a wall of text, and I’m back to reading, and trying to remember which tab I meant to check.
I built Pintail for the other half of that conversation: hearing the answer. A short summary, the project name, and a small companion on my Mac. The duck was not in the original requirements. The duck stayed.