sarvam-mcp 0.1.0

Official Sarvam AI MCP server — STT, TTS, Translate, Transliterate, LLM, Vision OCR for Indic languages

sarvam-mcp

Official Sarvam MCP server. Exposes every public Sarvam API — STT, TTS, Translate, Transliterate, Language ID, Text Analytics, LLM (Sarvam-M / 30B / 105B), Vision Document Intelligence, Pronunciation Dictionaries — as first-class MCP tools so any MCP-aware client (Claude Desktop, Claude Code, Cursor, Windsurf, Zed) can call Sarvam with zero boilerplate.

Quickstart

The fastest way is the one-line installer at mcp.sarvam.ai — auto-detects your MCP clients and wires them up:

curl -fsSL https://mcp.sarvam.ai/install | bash

Or install manually:

pip install sarvam-mcp        # or:  uvx sarvam-mcp

Drop this JSON into your MCP client (same shape works in Cursor / Claude Desktop / Claude Code / Windsurf / Zed):

{
  "mcpServers": {
    "sarvam": {
      "command": "uvx",
      "args": ["sarvam-mcp"]
    }
  }
}

No API key required up front. The server starts with auth deferred and prompts you for the key on the first tool call via MCP elicitation (Cursor / Claude Desktop will show a popup). The message links to Key management — open it, copy an API key, and paste. The key gets saved to ~/.sarvam/credentials (mode 0600) so subsequent runs don't ask.

If your MCP client doesn't support elicitation, or you'd rather set the key ahead of time (easiest first):

# A) Env var in the MCP client config (no terminal in many IDEs) — add next to
#    "args" for the sarvam server:  "env": { "SARVAM_API_KEY": "sk_..." }

# B) Interactive setup (headless / when you prefer the terminal)
sarvam-mcp init

# C) Advanced: create ~/.sarvam/credentials yourself (avoid echoing a real key in shell history)
mkdir -p ~/.sarvam && printf 'api_key = sk_...\n' > ~/.sarvam/credentials && chmod 600 ~/.sarvam/credentials

Per-client paths

• Claude Desktop — ~/Library/Application Support/Claude/claude_desktop_config.json
• Claude Code — claude mcp add sarvam -- uvx sarvam-mcp
• Cursor — ~/.cursor/mcp.json
• Windsurf — Cascade settings → MCP servers
• Zed — settings.json → context_servers

Tools

All defaults below reflect the latest non-deprecated models live as of 2026-04-27.

Tool	What it does	Default model	Other accepted
sarvam_stt_transcribe	Audio file → transcript (5 modes: transcribe, translate, verbatim, translit, codemix)	saaras:v3	saarika:v2.5 (legacy)
sarvam_stt_translate	Audio → English text (DEPRECATED — use stt_transcribe with mode=translate)	saaras:v2.5	—
sarvam_stt_batch_submit	Long-audio job init (Azure SAS)	saaras:v3	saarika:v2.5 (legacy)
sarvam_stt_batch_status	Long-audio job poll	—	—
sarvam_tts_speak	Text → audio file	bulbul:v3 (speaker priya)	bulbul:v3-beta, bulbul:v2
sarvam_tts_stream	Text → streamed audio	bulbul:v3	bulbul:v2
sarvam_translate	Cross-language text translate	mayura:v1	sarvam-translate:v1 (22 langs)
sarvam_transliterate	Script conversion	—	—
sarvam_identify_language	Language + script detect (11 languages)	—	—
sarvam_text_analytics	Typed Q&A over text	—	—
sarvam_llm_complete	Chat completions	sarvam-30b	sarvam-105b, sarvam-m (legacy)
sarvam_vision_extract	Document Intelligence (job-based pipeline)	Sarvam Vision (3B VLM)	—
sarvam_vision_job_status	Poll Document Intelligence job status	—	—
sarvam_pronunciation_list	List pronunciation dictionaries	—	—
sarvam_pronunciation_get	Get a pronunciation dictionary	—	—
sarvam_pronunciation_create	Create a pronunciation dictionary	—	—
sarvam_pronunciation_delete	Delete a pronunciation dictionary	—	—

Configuration

Env var	Default	Description
SARVAM_API_KEY	—	Required. API key. Falls back to ~/.sarvam/credentials.
SARVAM_API_REGION	in	Data residency region.
SARVAM_API_BASE_URL	https://api.sarvam.ai	Override for testing/staging.
SARVAM_MCP_BASE_PATH	~/Desktop	Where audio/document files land in files mode.
SARVAM_AUDIO_OUTPUT_MODE	files	files | resources | both.

~/.sarvam/credentials format:

api_key = sk_...
region = in

Two namespaces

The server exposes 27 tools across two clean namespaces:

• sarvam_tools_* — runtime tools. Call Sarvam APIs at runtime to do things (transcribe audio, generate speech, translate text, ask Sarvam-M, run composite voice/dub/localize/recall workflows). 16 tools.
• sarvam_code_* — builder tools. Help an agent write code that uses Sarvam: search docs, look up endpoint shapes, list supported languages and speakers, validate request bodies, recommend models, fetch tested code snippets, scaffold starter projects (simple-tts-cli, python-voice-bot, nextjs-translator). 11 tools — no API key needed for most of them.

"Translate this paragraph to Hindi." → sarvam_tools_translate invokes Sarvam.

"Build me an Indic translator app in Next.js." → sarvam_code_scaffold writes a working starter project to disk; sarvam_code_snippet provides tested glue code.

Companion repo

The install website at mcp.sarvam.ai lives in sarvamai/sarvam-mcp-website (Next.js + Tatva, deployed on Sarvam k8s). This repo is just the Python package.

Development

uv venv && source .venv/bin/activate
uv pip install -e ".[dev]"
pytest -q                              # 51 tests
mcp dev src/sarvam_mcp/server.py       # MCP Inspector