sarvamaisarvamai/skills/sarvam-mcp/SKILL.md

sarvam-mcp

Skill72 starsChanged 46 days agoApache-2.0View on GitHub

Run Indic speech and language tasks through the Sarvam MCP server: translate text, transcribe or dub audio, TTS, localize i18n files, LLM complete, document parse / extract / classify / split, or fetch Sarvam API docs/snippets while coding. Use this skill when the user wants Sarvam to act in chat, when sarvam_tools_* / sarvam_code_* are available, or when setting up sarvam-mcp in Cursor, Claude Code, Claude Desktop, Windsurf, Zed, Codex, Gemini CLI, VS Code, Cline, Continue, or LM Studio — even if they only say "translate this to Hindi" or "make it speak." Do not use for LiveKit/Pipecat agent code (voice-agents) or SDK-only coding when MCP is not installed (translate/text, translate/document, speech-to-text, text-to-speech, chat).

Reads credentials

In this repository7 other files it gives its agents.
- chatSkill
- doc-aiSkill
- dubbingSkill
- speech-to-textSkill
- text-to-speechSkill
- vibe-codingSkill
- voice-agentsSkill

The file

We split this skill into the rules it gives an agent

4 rules. About 0 are common, 1 is distinctive, and 3 are very specific to this project.

≈0
Common · estimate
Many files like this have a rule of this kind.
≈1
Distinctive · estimate
This author's own choices. Worth a look.
3
Very specific
Tied to their project's names, files or tools. Probably skip.

Common
Distinctive
Very specific

Very specific:
Never use sarvam_code_* for a live translate/TTS/STT request.
line 44
Very specific:
Call again with sk_... → persists to ~/.sarvam/credentials
line 111
Distinctive:
Read references/install.md and help the user connect the server.
line 160
Very specific:
If they only need code: use sibling skills translate/text, translate/document, speech-to-text, text-to-speech, chat, or …
line 161

≈ “Common” is an estimate: at least 1 in 5 of the 350,495 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.

Files

SKILL.md

Sarvam MCP

Drive sarvam-mcp. Prefer MCP tools over inventing HTTP/SDK calls for live work.

Procedure

Copy and track:

- [ ] 1. MCP available? If no → read references/install.md (or fall back to SDK skills)
- [ ] 2. Namespace: live action → sarvam_tools_*; write/lookup code → sarvam_code_*
- [ ] 3. Prefer a composite if the task spans STT/translate/TTS/LLM
- [ ] 4. Call the tool (absolute paths; BCP-47 codes)
- [ ] 5. On auth error → sarvam_tools_set_api_key; on param error → fix and retry
- [ ] 6. Return the result path/text; mention observability only if debugging

1. Namespace (mandatory)

User wants…	Namespace	Example
Sarvam to do it now	`sarvam_tools_*`	"Translate this to Tamil"
Help writing code / API facts	`sarvam_code_*`	"How do I call TTS from Python?"

Never use sarvam_code_* for a live translate/TTS/STT request. Never burn sarvam_tools_* credits just to draft code unless the user asked for a live demo.

2. Pick the tool (defaults)

Composites first — do not hand-chain STT → translate → TTS:

Task	Tool
Spoken reply to audio	`sarvam_tools_voice`
Dub into another Indic language	`sarvam_tools_dub`
Localize JSON/YAML/PO/etc.	`sarvam_tools_localize`
Q&A / summary over audio	`sarvam_tools_recall`

Atomic runtime (use when no composite fits):

Task	Tool
Transcribe	`sarvam_tools_stt_transcribe`
Speech → English	`sarvam_tools_stt_translate`
Audio >~30s / diarization	`sarvam_tools_stt_batch_submit` → `_stt_batch_status`
Speak text	`sarvam_tools_tts_speak`
Translate text	`sarvam_tools_translate`
Transliterate / LID / analytics	`sarvam_tools_transliterate` / `_identify_language` / `_text_analytics`
Chat complete	`sarvam_tools_llm_complete`
Document → text / HTML / JSON	`sarvam_tools_vision_parse`
Document fields (schema)	`sarvam_tools_vision_extract`
Label a document or pages	`sarvam_tools_vision_classify`
Split a multi-document PDF	`sarvam_tools_vision_split`
Resume / check a document job	`sarvam_tools_vision_job_status`
Pronunciation dicts	`sarvam_tools_pronunciation_*`
Set / rotate API key	`sarvam_tools_set_api_key`

Build-time (coding help):

Task	Tool
Unsure which model/lang	`sarvam_code_recommend_model`
Endpoint shape	`sarvam_code_api_reference`
Snippet	`sarvam_code_snippet`
Speakers / languages	`sarvam_code_speakers` / `_languages`
Validate draft body	`sarvam_code_validate_request`
Docs search / pricing	`sarvam_code_search_docs` / `_pricing`

Coding flow: recommend_model → snippet or api_reference → validate_request.

For parameters and env knobs, read references/tools.md.

3. Defaults (use unless user overrides)

Setting	Default
STT model	`saaras:v3`
TTS model / speaker	SDK code: `bulbul:v4-flash` / persona ID (`aparna_hi_customer`). Live `tts_speak` / `tts_stream` / `voice` / `dub` (sarvam-mcp ≥ 0.2.11): `model` / `tts_model` = `bulbul:v4-flash` + persona ID; omit the speaker for the `shubh_enhi_ads` default. On older servers the schema enums only `bulbul:v3` + short names (`priya`) — check the connected schema and do not send a v4 persona ID unless it lists `bulbul:v4-flash`.
LLM	`sarvam-30b` (use `sarvam-105b` for hard reasoning)
Translate model	`mayura:v1` (switch to `sarvam-translate:v1` for broader Indic coverage)
Audio path	Absolute local path
Language codes	BCP-47 (`hi-IN`, `ta-IN`, `od-IN` not `or-IN`)

4. Auth

Key sources (env wins): SARVAM_API_KEY, or ~/.sarvam/credentials (api_key = sk_...).

Dashboard: https://dashboard.sarvam.ai/key-management

On auth failure:

sarvam_tools_set_api_key with empty api_key → instructions + link
Call again with sk_... → persists to ~/.sarvam/credentials

Wire auth is api-subscription-key — never invent Authorization: Bearer.

Examples

Live translate

User: "Translate 'Good morning' to Hindi."

→ sarvam_tools_translate with input, source_language_code=en-IN, target_language_code=hi-IN. Return translated_text.

Live TTS

User: "Say नमस्ते in Hindi."

→ sarvam_tools_tts_speak with native-script text, target_language_code=hi-IN. Prefer model=bulbul:v4-flash with a persona ID (aparna_hi_customer) when the tool schema lists it (sarvam-mcp ≥ 0.2.11); on older servers use speaker=priya on bulbul:v3. Return the audio path.

Dub

User: "Dub this clip into Tamil" + path.

→ sarvam_tools_dub with audio_path, target_language_code=ta-IN — not STT + translate + TTS separately.

Code help

User: "Show a Python TTS example."

→ sarvam_code_snippet with api=tts, language=python (not sarvam_tools_tts_speak unless they want audio now).

Gotchas

Mistake agents make	Correct behavior
Invent `curl`/SDK for a live ask	Call `sarvam_tools_*` when MCP is connected
Hand-chain STT→translate→TTS	Use `voice` / `dub` / `localize` / `recall`
REST STT on long files	Use `stt_batch_*` above ~30s
TTS target outside ~11 langs	Check with `sarvam_code_languages` (`api=tts`); STT has ~23
v2 speaker on Bulbul v3	Use `priya`/`shubh` or `sarvam_code_speakers`
v3 short name on `bulbul:v4-flash` (or a v4 persona on v3)	Speaker/model mismatch is rejected. v4-flash takes persona IDs only (`aparna_hi_customer`); v3 takes `shubh` / `priya`. Check `sarvam_code_speakers` with `model=`.
`pitch` / `loudness`	`pitch` −0.5–0.5, `loudness` 0.1–2.5, `pace` 0.5–2.0; out-of-range values are rejected. Accepted on both v3 and v4-flash (live-verified 2026-10-06).
Romanized Indic for TTS	Prefer native script
`output_script` on `sarvam-translate:v1`	`mayura:v1` only
Relative audio paths	Prefer absolute paths; or `audio_base64`/`audio_url` + `filename`
Assume output cwd	Files go under `SARVAM_MCP_BASE_PATH` (default `~/Desktop`) unless tool returns another path
Stale tool names without `_tools`/`_code`	Match the connected server's tool list

When MCP is missing

Read references/install.md and help the user connect the server.
If they only need code: use sibling skills translate/text, translate/document, speech-to-text, text-to-speech, chat, or https://docs.sarvam.ai/llms.txt.

References (load on demand)

File	Read when
references/install.md	MCP not connected, setup/troubleshoot, or user names a client
references/tools.md	Choosing parameters, env vars, or an uncommon tool

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

It worked
Partly
It didn't

What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

Check your setup, free