Setu

Quickstart

Install & run in 3 steps

Defaults to mock mode — deterministic fixtures, no API key, no credits burned.

bash — install

# from PyPI
pip install setu-mcp

# or run without installing
uvx setu-mcp

# or from source
pip install git+https://github.com/manasasb2000/setu-mcp.git

bash — configure (optional)

# mock is the default — nothing to set
export SETU_MODE=mock

# for real Sarvam calls, switch to live:
# export SETU_MODE=live
# export SARVAM_API_KEY=sk_...   (dashboard.sarvam.ai)

bash — run

# stdio — for Claude Desktop / Cursor / local agents
setu

# or an HTTP endpoint at http://127.0.0.1:8000/mcp
setu --transport streamable-http

Connect to Claude Desktop

Install the package

pip install setu-mcp into a Python 3.11+ environment.

Add it to the config

Claude Desktop → Settings → Developer → Edit Config, then add the block below.

Restart & use it

Reopen Claude Desktop, click the 🔌 tools icon, and ask: "Use sarvam_transcribe on this Hindi audio."

claude_desktop_config.json

{
  "mcpServers": {
    "setu": {
      "command": "setu",
      "env": { "SETU_MODE": "mock" }
    }
  }
}

The tool surface

8 typed tools, one install

Tool	Model	What it does
setu_ping	—	Health check — confirms the server is wired up and reports mode.
sarvam_transcribe	Saaras v3	Speech → text, including code-mixed (Hinglish) audio.
sarvam_speak	Bulbul v3	Text → natural speech (base64 WAV), 11 languages, 30+ voices.
sarvam_translate	Mayura v1	Translate between English and 22 Indian languages.
sarvam_transliterate	Mayura	Convert script while keeping pronunciation (नमस्ते ↔ namaste).
sarvam_identify_language	text-lid	Detect the language + script of a piece of text.
sarvam_chat	Sarvam-105B / 30B	Chat / reasoning with OpenAI-compatible tool-calling.
sarvam_parse_document	Parse	PDF / image → structured JSON fields (KYC-ready).

Every result also carries cost_inr, latency_ms, and a request_id.

How it works

Every call: validated, retried, priced

MCP client (Claude Desktop / Cursor / LangGraph) │ stdio OR streamable-HTTP / SSE ▼ ┌──────────────────── SETU MCP SERVER ────────────────────┐ │ FastMCP router → typed tool layer (Pydantic in/out) │ │ validate → rate-limit → retry/backoff → │ │ call Sarvam → normalize errors → telemetry │ │ │ │ │ │ ┌───────┴──────┐ ┌──────┴──────┐ │ │ │ MODE: live │ │ OTel spans │ │ │ │ / mock │ │ cost+latency│ │ │ └───────┬──────┘ └─────────────┘ │ └──────────────────┼──────────────────────────────────────┘ ▼ Sarvam APIs — Saaras · Bulbul · Mayura · Sarvam-30B/105B · Parse

Reliability

Exponential backoff + jitter on 5xx / network errors; never retries 4xx except 429.
Client-side token-bucket rate limiter; respects Retry-After on 429.
Typed errors surfaced to the agent on final failure.

Observability

Every call wrapped in an OpenTelemetry span.
Attributes: model, endpoint, latency_ms, input units, computed cost_inr.
Maintained INR pricing table per model / endpoint.

Two transports

stdio for local IDE / desktop agents.
streamable-http / SSE for hosted, remote use.
Both served from one FastMCP server object.

Mock vs live

SETU_MODE=mock|live switch.
Mock serves deterministic fixtures — CI & contributors never burn credits.
Demos work fully offline.

Deploy anywhere

Run it as a container

A multi-stage, non-root image serves the HTTP transport out of the box.

bash — docker

# build and run the HTTP server on :8000/mcp
docker build -t setu-mcp .
docker run --rm -p 8000:8000 setu-mcp

# or one command with compose
docker compose up --build

# live mode
docker run --rm -p 8000:8000 -e SETU_MODE=live -e SARVAM_API_KEY=sk_... setu-mcp

Sarvam's Indic AI stack, as agent tools