Run it yourself.

Five commands, three ports and a few keys. The Gnani and language model keys switch on speech and replies; the Vobiz keys switch on the phone line.

Five commands.

git clone https://github.com/ayushcody/vakya.git && cd vakya

Get the code.

2
```
./vakya setup
```
Checks Python 3.10+ and Node 20+, builds backend/.venv, installs the web packages and creates backend/.env.
3
```
nano backend/.env
```
Add GNANI_API_KEY and the LLM_ and EMBEDDINGS_ settings, and the Vobiz keys for phone calls (see the table below).
4
```
./vakya start
```
Starts the API on port 8020 and the web app on port 3010 in the background.
5
```
./vakya doctor
```
Checks the machine and the running app and prints a fix for every problem.

The repository is private until the challenge closes, so the clone step needs access: ask the team. Then open http://localhost:3010/desk/try and make a call from the browser.

Three ports.

3010: The web app: this site, the desk at /desk and /verify.
8020: The API. Keep it on your machine; it is never put on the internet.
8021: The phone-only public port. A tunnel points Vobiz at it, and it serves only the call stream, the answer document, the hangup callback, the missed-call link and the web-call page.

Keys, and what you lose without them.

Environment settings in backend/.env, what each gives and what happens without it
Setting	Needed	It gives you	Without it
GNANI_API_KEY	Required	Gnani Prisma v3 to listen, Timbre to speak, and vachana-vc-v1 voice cloning.	Vakya cannot hear or speak. There is no second speech model: the desk, schedules and bookings still open, but no call works.
LLM_API_KEY, LLM_BASE_URL, LLM_MODEL	Required	A hosted language model, set in .env, writes the front desk's replies and chooses its tools.	No replies: the agent answers every message with a fixed line.
EMBEDDINGS_API_KEY, EMBEDDINGS_BASE_URL, EMBEDDINGS_MODEL	Recommended	Hosted embeddings search the business's documents.	Document search falls back to keyword search.
VOBIZ_AUTH_ID, VOBIZ_AUTH_TOKEN, VOBIZ_CALLER_ID	For phone calls	Real phone calls through a Vobiz number. ./vakya start opens a tunnel so Vobiz can reach your computer. For missed-call callbacks, make a Vobiz XML Application whose Answer URL is the link from Front desk, Phone number, and attach a number to it; built and tested with fakes, not yet tried on a live number.	Everything else works, including the browser web call. A trial account only calls verified numbers.
GOOGLE_CALENDAR_ID, GOOGLE_SERVICE_ACCOUNT_FILE	For the calendar	Bookings sync to a Google Calendar shared with the service account.	Bookings are kept in Vakya's own list.

Keys live only in backend/.env on your computer. The Gnani paths are integrated but have not run against the live API yet; the first thing to do with a key is the smoke test below.

Check that it works.

```
backend/.venv/bin/python backend/bench/gnani_smoke.py
```
First contact with Gnani's API with your key: lists voices, runs Prisma on a test clip, renders Timbre in two formats. Prints every reply shape and never prints the key.
```
./vakya stress --conditions all --provider gnani --keyterms
```
The acoustic-stress report card on Prisma: the same 24 invented booking calls, clean, 8 kHz, noisy and through a phone codec, with and without keyterms. First run 6 Nov 2026, published on /report-card.
```
./vakya test
```
The backend and frontend test suites.

Which machine.

Tried on: Apple M4 with 16 GB, macOS. Linux, Windows (a vakya.ps1 script exists) and Macs with 8 GB are untested.

Every model is reached over the network (Gnani for speech, a hosted language model for replies), so nothing is downloaded and the laptop only has to run Python and Node. A call needs the internet the whole time, and a laptop that stays awake if the phone line is served from it.

Stuck? The FAQ lists what is not ready yet.