sarvam-mcp

Runtime tool — calls Sarvam API now. For code-writing help, use sarvam_code_* tools. Transcribe an audio file in any of 23 Indian languages using Saaras v3. Saaras v3 supports multiple output modes via the `mode` parameter: • `transcribe` (default) — standard transcription in the original language • `translate` — speech from any Indic language directly to English text • `verbatim` — exact word-for-word, no normalization, filler words preserved • `translit` — romanization to Latin/Roman script • `codemix` — English words in English, Indic words in native script The default `language_code='unknown'` auto-detects, but specifying the language (e.g. `hi-IN`, `ta-IN`) gives better accuracy. For very long files (>30s), prefer `sarvam_stt_batch_submit`.

Check for updates and upgrade sarvam-mcp to the latest version. Call this when the user asks to update, upgrade, or check for newer versions of the sarvam MCP server. Set ``confirm_upgrade=True`` to actually perform the upgrade. After upgrading, restart the MCP server for changes to take effect.

Runtime tool — calls Sarvam API now. Poll the status of a batch transcription job. Returns the transcript once `job_state == 'Completed'`.

Set or update your Sarvam API key. Call this if you get an authentication error or to configure your key for the first time. If called without a key, returns instructions on where to get one. If called with a key, saves it globally so you never need to set it again.

Runtime tool — calls Sarvam API now. For code-writing help, use sarvam_code_* tools. Transcribe a long audio file (>30 s) using the batch job pipeline. Runs the full flow automatically: create job → upload audio to Azure Blob → start processing → poll until complete → return transcript. Supports diarization, timestamps, and all Saaras v3 output modes.

Runtime tool — calls Sarvam API now. For code-writing help, use sarvam_code_* tools. DEPRECATED: Use `sarvam_tools_stt_transcribe` with `mode='translate'` instead. Transcribe an Indic-language audio file directly into English text using the legacy `/speech-to-text-translate` endpoint. This endpoint will be removed in a future version.

Runtime tool — calls Sarvam API now. For code-writing help, use sarvam_code_* tools. Generate speech from text (model bulbul:v3). 11 Indic languages. Speaker hints (v3 voice roster): • `priya` / `neha` / `pooja` — warm friendly female (default `priya`) • `aditya` / `rahul` / `kabir` — professional male • `shreya` / `kavya` / `ritu` — calm news-anchor female • `vijay` / `gokul` / `anand` — mature authoritative male • `tanya` / `suhani` / `niharika` — young energetic female The audio file is written under SARVAM_MCP_BASE_PATH (default ~/Desktop).

Runtime tool — calls Sarvam API now. For code-writing help, use sarvam_code_* tools. Detect the language and script of input text. Returns BCP-47 language code (e.g. 'hi-IN') and script code (e.g. 'Devanagari'). Useful as a pre-step before TTS or translate to pick the right language args automatically.

Runtime tool — calls Sarvam API now. Poll the status of an existing Document Intelligence job. Returns the current job state and page metrics. Once state is 'Completed', the output can be downloaded from the output URL.

Runtime tool — calls Sarvam API now. Delete a pronunciation dictionary by its ID.

Runtime tool — calls Sarvam API now. For code-writing help, use sarvam_code_* tools. Streaming variant of sarvam_tts_speak using the TTS WebSocket. Audio is streamed to disk in the background; the tool returns a sarvam:// resource URI immediately.

Runtime tool — calls Sarvam API now. For code-writing help, use sarvam_code_* tools. Translate text between English and Indian languages. Two backing models: - `mayura:v1` — 11 languages, supports formal / colloquial / code-mixed modes, script + numeral controls. Best for short, stylized translations. - `sarvam-translate:v1` — all 22 scheduled languages + English, formal only. Best when you need broad coverage.

Runtime tool — calls Sarvam API now. For code-writing help, use sarvam_code_* tools. Convert text between scripts without translating meaning. Examples: Devanagari → Roman ('नमस्ते' → 'namaste'), Tamil script → Latin, etc. Use `spoken_form=True` to expand digits and abbreviations into spoken form.

Runtime tool — calls Sarvam API now. For code-writing help, use sarvam_code_* tools. Run deep analysis on a piece of text by passing a list of typed questions. Each question needs `id`, `text`, and `type` (`boolean` | `enum` | `short answer` | `long answer` | `number`). Returns one answer per question, grounded in the text. Good for entity extraction, classification, and structured Q&A without writing prompt boilerplate.

Runtime tool — calls Sarvam API now. List all pronunciation dictionary IDs owned by the authenticated user. Returns dictionary_count and a list of dictionary IDs.

Build-time tool — helps write code that uses Sarvam. For runtime actions, use sarvam_tools_* instead. Return the authoritative request/response shape for a Sarvam API endpoint — parameters, defaults, content type, auth header, common gotchas. Faster + more accurate than free-text doc search when the agent already knows which endpoint to use.

Build-time tool — helps write code that uses Sarvam. For runtime actions, use sarvam_tools_* instead. Return a tested, copy-paste-ready code snippet for calling a Sarvam API. The agent should use this as the starting point when adding Sarvam to an existing codebase. Snippets are verified live against api.sarvam.ai with current model versions; they assume `SARVAM_API_KEY` is set in the env.

Runtime tool — calls Sarvam API now. For code-writing help, use sarvam_code_* tools. Generate chat completions with Sarvam's Indic-tuned LLMs. Models: • `sarvam-30b` (default) — MoE, 2.4B active, balanced quality + cost • `sarvam-105b` — MoE flagship, best reasoning + tool use Both support 23 Indic languages with native, romanized, and code-mixed styles. OpenAI-compatible message format.

Runtime tool — calls Sarvam API now. Retrieve a specific pronunciation dictionary by its ID. Returns the dictionary entries (word → pronunciation mappings).

Build-time tool — helps write code that uses Sarvam. For runtime actions, use sarvam_tools_* instead. List supported BCP-47 language codes for a given Sarvam API. Critical because coverage varies by API: STT supports 23 langs, TTS supports 11, etc. Returns code + display name + script.

Runtime tool — calls Sarvam API now. For code-writing help, use sarvam_code_* tools. Extract text + structure from a document or image using Sarvam Vision (Document Intelligence). Supports 23 Indian languages with table preservation. Outputs markdown (default), HTML, or JSON. This runs the full async pipeline: create job → upload file → start → poll until complete. Max 10 pages per document. Returns the output download URL (presigned) and job metadata. The output is delivered as a ZIP file containing the chosen format plus a JSON file with page-level data.

Runtime tool — calls Sarvam API now. Create a new pronunciation dictionary with word → pronunciation mappings. These dictionaries can be referenced in TTS calls to control how specific words are spoken. Max 100 words per dictionary, 10 dictionaries per user.

Build-time tool — helps write code that uses Sarvam. For runtime actions, use sarvam_tools_* instead. List TTS speakers compatible with a given model tag. The v3 roster has 38 voices. Returns each speaker with a brief tone hint where available.

Build-time tool — helps write code that uses Sarvam. For runtime actions, use sarvam_tools_* instead. Validate a draft Sarvam API request body against the canonical schema BEFORE the dev sends it. Catches: invalid model names, speaker/model mismatches, deprecated values, unsupported language codes, missing required fields. Returns a list of issues with fix suggestions.

Build-time tool — helps write code that uses Sarvam. For runtime actions, use sarvam_tools_* instead. Current pricing structure for a Sarvam model (or all models if omitted). Returns billing unit + rate tier. NOTE: this is a high-level structure; exact per-unit rates depend on your plan — always confirm at https://dashboard.sarvam.ai before quoting.

Runtime tool — calls Sarvam API now. For code-writing help, use sarvam_code_* tools. End-to-end voice agent loop: transcribe an Indic audio file, send the transcript to the Indic-tuned LLM for a reply, then synthesize the reply to audio. Returns: transcript, reply text, output audio file path. Default reply language follows the detected input language; set `reply_language` to force a specific output.

Build-time tool — helps write code that uses Sarvam. For runtime actions, use sarvam_tools_* instead. Given a plain-English description of what the dev wants to build, recommend the right Sarvam model + language code + any other key params. Useful as the first step when starting a new Sarvam-using app. Uses heuristics (no LLM call), so it's fast and works without an API key.

Runtime tool — calls Sarvam API now. For code-writing help, use sarvam_code_* tools. Dub an Indic audio file into another Indic language. Pipeline: STT → Mayura or Sarvam-Translate → TTS. Returns the original transcript, translated transcript, and a path to the newly-dubbed WAV file.

Runtime tool — calls Sarvam API now. For code-writing help, use sarvam_code_* tools. Translate a JSON or `key=value` string-table file into another language. Walks all string values (nested JSON supported), runs each through Sarvam-Translate, and writes a sibling file with the language code as suffix (e.g. `messages.en.json` → `messages.hi.json`). For repos with many files, call this tool once per locale file. Skips non-string leaves (numbers, booleans, etc.) untouched.

Runtime tool — calls Sarvam API now. For code-writing help, use sarvam_code_* tools. Answer a natural-language question grounded in a set of files. Audio files are transcribed; text/markdown files are read directly. Everything is concatenated (capped at 24000 chars) and passed to the LLM with a grounded-QA system prompt. Useful for quickly querying meeting recordings, Indic-language notes, or mixed-format research folders.