Full-stack AI for Bharat
Documentation • Get API Key • Agent Skills Spec
LLMs have fixed knowledge from their training cutoff. Sarvam AI's SDK has unique patterns that differ from standard conventions — method names that break expectations (client.text.translate() not client.translate.translate()), parameters that silently fail (output_script on sarvam-translate), and response quirks (content being None when reasoning consumes the token budget).
These skills bridge that gap. Each one gives AI coding assistants the exact SDK signatures and gotchas they need to generate correct Sarvam AI code, then routes to llms.txt for detailed documentation.
For in-chat Sarvam actions (translate this, speak that, dub audio) across Cursor, Claude Code, and other MCP clients, install sarvam-mcp and use the sarvam-mcp skill.
The vibe-coding skill is stack- and vendor-agnostic—it is for anyone building with an agent. sarvam-mcp teaches agents to drive the Sarvam MCP server in any harness. The other skills document Sarvam SDK APIs for writing code.
| Skill | Description |
|---|---|
| sarvam-mcp | MCP (any harness) — live sarvam_tools_* vs build-time sarvam_code_*, composites, auth, install. Use for in-chat Sarvam actions. |
| chat | SDK — Sarvam-105B/30B completions, streaming, reasoning, content=None gotcha. |
| speech-to-text | SDK — Saaras v3 REST, Batch + diarization, WebSocket streaming. |
| text-to-speech | SDK — Bulbul v3 REST/stream/WebSocket, pronunciation dicts, v3 param traps. |
| translate | SDK — Mayura / Sarvam-Translate signatures and silent failures. |
| voice-agents | SDK — LiveKit / Pipecat real-time voice agents. |
| vibe-coding | Vendor-neutral agent habits (slice → verify → iterate). Pair with a domain skill for APIs. |
# Install all skills
npx skills add sarvamai/skills
# Install a specific skill
npx skills add sarvamai/skills --skill sarvam-mcp
npx skills add sarvamai/skills --skill chat
npx skills add sarvamai/skills --skill vibe-coding
# Browse skills interactively
npx skills add sarvamai/skills --list# Setup (SDK skills)
export SARVAM_API_KEY="your-api-key" # get at dashboard.sarvam.ai
pip install sarvamai # Python
npm install sarvamai # JavaScript/TypeScript
# Setup (MCP skill — live tools in-chat)
# uvx sarvam-mcp — see sarvam-mcp/references/install.md for each clientWorks with Cursor, Claude Code, Windsurf, and any agent that supports the Agent Skills specification. The MCP skill also covers Claude Desktop, Zed, Codex, Gemini CLI, VS Code, Cline, Continue, and LM Studio.
sarvam-mcp/SKILL.md ← MCP routing (tools_* vs code_*) for any harness
SDK skill/SKILL.md ← SDK signatures + gotchas (what agents get wrong)
│
▼
llms.txt / MCP code_* ← Always-fresh docs index or live API reference tools
│
▼
Full API docs, OpenAPI spec, cookbooks, voice catalog, streaming protocols...
sarvam-mcp teaches agents when to call live MCP tools vs build-time helpers, and how to install the server in each client. The other Sarvam skills are a lean correction layer with both Python and JavaScript/TypeScript SDK snippets — only what AI agents get wrong when generating Sarvam AI code:
- SDK call signatures that differ from conventions (e.g., no
.create()on chat) - Parameters that silently fail (e.g.,
output_scriptignored on sarvam-translate) - Parameters that error (e.g.,
pitch/loudnessreturns 400 on Bulbul v3) - Non-trivial SDK patterns (e.g., Batch API job chain, WebSocket async connect)
For everything else — full parameter tables, voice catalogs, language codes, rate limits, cookbook examples — the skill points to llms.txt, which is always up to date.
Apache-2.0