█▀▄ █ ▀█▀ █ █ █▀▄ █ ▀█▀ █ █
█▀  █  █  █▄█ █▀  █  █  █▄█

█▀▄ █ ▀█▀ █ █ █▀▄ █ ▀█▀ █ █
█▀  █  █  █▄█ █▀  █  █  █▄█

Delegate work to AI agents by voice, from your phone. Keep track of it while you drive.

pitupitu is a foreman for your coding agents. Say what you want done; it starts a Claude Code or Codex session on your own machine, watches it, and tells you when it finishes or needs you. Nothing to type. Open source.

start free — 30 minutes how it works

phone — listening

▁▂▃▂▁▂▄▆▄▂▁▁▂▃▅▇▅▃▂▁▂▃▂▁▁▃▅▃▁▂▄▂▁▁▂▃▂▁▂▅▇▅▂▁

mic open · 24 kHz · streaming

pitupitu serve — transcript

youstart a task in the api repo: add rate limiting to the login endpoint

pitupituon it. I'll tell you when it's done.

— 6 minutes later, between two of your sentences —

pitupituthe login task finished, tests green.

youwhat did it change?

pitupitua token bucket per account on /login, a 429 with retry-after, and two tests.

How it works

step 1
Install next to herdr
```
curl -fsSL https://pitupitu.chat/install.sh | sh
```
One binary on the machine where your agents already run. herdr is the terminal multiplexer that keeps every agent session in view.
step 2
Link it in the terminal
```
pitupitu setup
```
Checks herdr and which agents are logged in, then prints a code. Open the link on any screen, approve the box, and the terminal is signed in. No keys to paste.
step 3
Talk from your phone
```
pitupitu serve
```
Open the page on your phone over your tailnet. Say what you want done. Interrupt it mid-sentence. It reports in the pauses, never over you.

Runs on your box, your subscriptions

The agents are yours. pitupitu drives the official claude and codex CLIs exactly as you would, so the work runs on the subscriptions you already pay for. We never proxy those tokens.

What we host is the part that needs a GPU: streaming speech in and out, and the small model that runs the conversation. Both are billed by the minute you talk, and both can be swapped for your own GPU or keys.

pitupitu.toml

[cloud]
api_key = "ptp_…"

[speech]
kind = "cloud"   # or "kyutai" on your own GPU

[model]
kind = "cloud"   # or any OpenAI-compatible endpoint

[places]
api = "~/code/api"

Pricing

Trial

30 minutes, no card

30 minutes, no card. Enough to find out whether it fits your day.

start

monthly

Starter

from $18/month

300 minutes included · $0.06/min

A month of daily commutes. Unused minutes roll over once.

Top-up

$10

150 extra minutes · $0.07/min

For a heavy week. Minutes never expire while your account exists.

top up

Minutes count while the microphone is connected. There is no unlimited plan: every minute runs on a GPU, and we would rather tell you that than hide it.

Open source

Everything that runs on your machine is open: the server, the phone client, the speech client and the Docker image for Kyutai's streaming models. Run it on a box with a GPU and you never need this site.

github.com/klarluft/pitupitu

Questions

? What do you store?: Nothing you say. Audio, transcripts and the conversation are relayed and never written down. We keep your email, a hash of each box's key and how many seconds each session lasted. That is the whole database.
? Does it work without your cloud?: Yes. With an NVIDIA GPU (about 8 GB of VRAM) you run Kyutai's streaming speech models yourself and point the foreman at any model endpoint, local or hosted. The hosted tier is the same software with the GPU on our side.
? Which agents?: Anything herdr can start: Claude Code, Codex, Gemini CLI, OpenCode and twenty-odd others. They run under your own logins and permission settings; pitupitu only types and watches.
? Does it do anything irreversible on its own?: No. Stopping a working session, pressing keys in an agent's dialog, anything that cannot be undone, needs a spoken confirmation that pitupitu reads back first. The check is in code, not in the model's hands.
? Why minutes and not tokens?: The expensive tokens are the agents', and they run on your subscription. What we pay for is GPU time while you talk, so that is what you pay for.