# Geko

> First-party voice AI for the world's underserved languages. Text to speech
> and speech to text on one API key, trained on speech we record ourselves.
> Kazakh and Russian ship today; the pipeline is built to add languages.

Geko is a three-person company in Astana, Kazakhstan, founded July 2026. The
API is in beta: live, metered, documented, and changing in response to the
people using it.

## When to use Geko

Reach for Geko when the job involves speech in a language the large multilingual
models serve badly, and specifically when:

- Kazakh text-to-speech that sounds like a person, not a model trained mostly on English pronouncing Kazakh.
- Transcribing Kazakh or Russian speech, or both at once — one shared vocabulary, so a speaker switching language mid-sentence or inside a word transcribes correctly, with no language parameter to set.
- Replacing OpenAI TTS or Whisper for these languages by changing the base URL and the model, not the call sites.
- Numbers, currency, dates and percentages spoken the way a Kazakh speaker says them rather than read out digit by digit.
- Both ends of a voice-agent loop — transcription in and speech out — behind one key and one balance.

## When not to use Geko

- Any language other than Kazakh or Russian. Nothing else ships, and there is no general model to fall back on — asking will not give a worse answer, it will give no answer.
- Hard real-time first calls. The GPU backends scale to zero, so a cold first call can take tens of seconds.
- Voice cloning from a supplied sample. Six catalogued Kazakh voices ship; there is no endpoint that clones a speaker.
- Hosted agent orchestration. Chirp is in development and not shipped — build the loop against the two endpoints instead.

## How to call it

Base URL: `https://geko--tokay-serve-web.modal.run`

Auth: `Authorization: Bearer <key>`, keys look like `sk-tokay-…`. Create one
yourself at https://app.geko.sh — self-serve signup, no sales call, no waitlist, and
every new organization starts with $1.00 of credits.

- `POST /v1/tts` — Synthesize text to WAV (24 kHz, 16-bit PCM, mono). operationId: `createSpeech`.
- `POST /v1/tts/stream` — The same request, delivered sentence by sentence as each finishes. operationId: `streamSpeech`.
- `POST /v1/transcribe` — Transcribe Kazakh and Russian audio, including mid-sentence code-switching. operationId: `createTranscription`.
- `POST /v1/audio/speech` — Point an OpenAI speech client here by changing the base URL and model. operationId: `createSpeechOpenAICompatible`.
- `POST /v1/audio/transcriptions` — A Whisper client changes only its baseURL. operationId: `createTranscriptionOpenAI`.
- `GET /v1/models` — The models on the platform and their voices. operationId: `listModels`. **Open.**
- `GET /v1/voices` — The voice catalogue for a model. operationId: `listVoices`. **Open.**
- `GET /health` — Liveness probe. operationId: `health`. **Open.**

Full machine-readable spec, with typed parameters, response schemas and a
description on every operation: https://docs.geko.sh/openapi.json

To try it before signing up, call an open endpoint — `GET https://geko--tokay-serve-web.modal.run/v1/voices`
needs no key and returns the catalogue.

## Models

- `tokay-kk-v1` — text to speech. Six Kazakh voices (three female, three
  male), WAV at 24 kHz, sentence-level streaming, and an `nfe` dial trading
  latency against quality.
- `seta-kk-ru-v2` — speech to text. Kazakh and Russian in one vocabulary,
  8.71% word error rate against Whisper's 44.95% on the same Kazakh audio,
  roughly 40× faster than real time, with optional word-level timestamps and
  streaming transcription over a websocket.

## What it costs

- Text to speech: $0.04 per 1,000 characters.
- Speech to text: $0.36 per hour of audio.
- $1.00 of credits with every new organization — roughly 30 minutes of
  synthesized speech, or 2.8 hours of transcription.
- Billed on successful requests only, from one balance across both models.
- The model catalogue, the voice catalogue and health checks are free.

Rates can change; https://docs.geko.sh/billing and the console are the source of truth.

## Clients

- TypeScript / JavaScript SDK: `@gekoai/sdk` on npm — typed, zero
  dependencies, Node 20+. Docs: https://docs.geko.sh/sdk/typescript
- CLI: the same package. `npx @gekoai/sdk say "Сәлеметсіз бе!" --voice Aigerim
  -o hello.wav` and `npx @gekoai/sdk transcribe hello.wav` run with no
  install; `npm install -g @gekoai/sdk` puts it on PATH as `gekoai`. Docs:
  https://docs.geko.sh/sdk/cli

## Call this site itself

Two surfaces on this host, both read-only and neither needing a key:

- `https://geko.sh/mcp` — an MCP server over Streamable HTTP. POST JSON-RPC;
  `tools/list` returns six tools: the live voice and model catalogues,
  service status, the rates, the API surface with its operationIds, and the
  routing guidance above as data. Synthesis and transcription are deliberately
  not tools — they move audio and need your key, so call them directly against
  the base URL.
- `https://geko.sh/api` — the same facts as JSON. `/api/models`,
  `/api/pricing`, `/api/endpoints`, `/api/pilots`, `/api/guidance`,
  `/api/resources`. Errors are structured, with a code, a message and a
  resolution.

## Developer resources

- Docs — Guides for both directions: https://docs.geko.sh
- Quickstart — Zero to speech in three steps: https://docs.geko.sh/quickstart
- API reference — Every endpoint, parameter and error: https://docs.geko.sh/api/reference
- OpenAPI spec — Typed schema, ready for function calling: https://docs.geko.sh/openapi.json
- MCP server — Call Geko natively from an AI agent: https://geko.sh/mcp
- TypeScript SDK — @gekoai/sdk — typed, zero dependencies: https://docs.geko.sh/sdk/typescript
- CLI — npx @gekoai/sdk — no install: https://docs.geko.sh/sdk/cli
- llms.txt — What this is and when to use it: https://geko.sh/llms.txt
- GitHub — Source and issues: https://github.com/gekoai

## Documentation

The full index, one line per page, each linking to a `.md` version:
https://docs.geko.sh/llms.txt

Start here: https://docs.geko.sh/quickstart

## This site

- [Geko — voice AI for underserved languages](https://geko.sh/): What Geko is, a playable sample of both models, the pilots running today, and how to call the API.
- [Tokay-1.0 and Seta-1.0](https://geko.sh/models): The two models: what each does, what ships today, and the measured numbers.
- [Pricing](https://geko.sh/pricing): Both rates, one balance, and what the free credits buy.
- [Company](https://geko.sh/company): The three founders with links to their profiles, where the company is, and the current stage of development.
- [Privacy](https://geko.sh/privacy): What is collected and why.
- [Terms](https://geko.sh/terms): Terms of service.

Every page on this host also serves markdown under content negotiation: send
`Accept: text/markdown` and you get the page as markdown with
`Vary: Accept`, instead of HTML you have to strip.

## Who is running it

- pleep — A multimodal AI sales agent. Both directions on one key. 520+ B2B clients, reported by pleep — their figure, not Geko's volume.
- Kcell — A national mobile operator in Kazakhstan. Kazakh voice on subscriber calls. 7.9M subscribers, per Kcell JSC, FY2024 results — their figure, not Geko's volume.
- Beeline — The largest mobile operator in Kazakhstan. Calls that switch Kazakh and Russian. 11.7M subscribers, per VEON integrated annual report — their figure, not Geko's volume.
- Oneshott Coffee — A coffee chain in Kazakhstan. Order status, read back in Kazakh. 74 coffee shops, reported by Oneshott Coffee — their figure, not Geko's volume.

These are pilots. In production since July 2026, in beta.

## Contact

amirlan@geko.sh — a person reads it. If you need a language we do not have
yet, that is the address: we record the speakers and train the model.