June 12, 2026API & SDK Update

RAIL Score MCP server is live, plus Python and JavaScript SDK 2.6

The RAIL Score MCP server is now live at mcp.responsibleailabs.ai, bringing responsible AI evaluation, compliance, agent safety, and DPDP tools directly into any MCP-compatible client. The Python (2.6.1) and JavaScript/TypeScript (2.6.0) SDKs add the DPDP namespace, Indian PII detection, and runtime configuration introspection. We have also published the RAIL Guard Benchmark on Hugging Face.

Two updates land together this month: the RAIL Score Model Context Protocol (MCP) server is now live, and the official Python and JavaScript/TypeScript SDKs have shipped version 2.6. Both bring the full RAIL Score evaluation surface, including the India DPDP toolkit, into the places developers and AI agents already work.

RAIL Score MCP server

The RAIL Score MCP server is a remote, stateless gateway that exposes the RAIL Score API to any MCP-compatible client, including agent frameworks and AI-powered editors. It runs in front of the same production API, so credits, tenancy, and rate limits are identical to direct REST access.

Endpoint:

https://mcp.responsibleailabs.ai/mcp

Authenticate with your RAIL API key via Authorization: Bearer rail_… or X-API-Key: rail_….

The server ships nine tools covering the full evaluation surface:

rail_evaluate

Score content across the eight RAIL dimensions in basic or deep mode, with optional per-dimension policy rules.

rail_check_compliance

Test outputs against one to five regulatory frameworks in a single call.

rail_detect_injection

Fast prompt-injection scan that returns a verdict only, never echoing the analyzed text.

rail_evaluate_tool_call

Pre-execution ALLOW / FLAG / BLOCK decision for agent tool calls.

rail_scan_tool_result

Detect PII in tool output, returning types and offsets with a locally redacted copy, never raw values.

rail_safe_regenerate

Iteratively evaluate and regenerate content until it meets a quality threshold.

rail_dpdp_scan, rail_dpdp_gate, rail_dpdp_compliance

India DPDP toolkit: scan for Indian PII, gate processing against a declared purpose, and emit or require compliance evidence.

The server is published to the official MCP registry as ai.responsibleailabs/rail-score and listed on Smithery. Verdicts are always returned as structured data, analyzed text is never echoed back, and detected PII is never returned in the clear.

SDK Updates (2.6)

Python SDK v2.6.1

pip install rail-score-sdk

Adds the DPDP namespace, Indian PII recognizers, and configuration introspection helpers to read available dimensions, frameworks, and account capabilities at runtime.

View on PyPI →

JavaScript / TypeScript SDK v2.6.0

npm install @responsible-ai-labs/rail-score

Brings the same DPDP namespace, mirrored Indian PII constants, and configuration introspection to type-safe JavaScript and TypeScript projects.

View on npm →

New benchmark: RAIL Guard

Alongside these releases we have published the RAIL Guard Benchmark on Hugging Face, an Apache 2.0 dataset for evaluating model safety across both content generation and agentic tool use.

RAIL Guard Benchmark

Two pools across six domains (healthcare, finance, legal, education, customer support, and code generation): 1,197 content-evaluation prompts at varying difficulty tiers, and 392 agent scenarios that test tool-call safety decisions. Every example is scored across the eight RAIL dimensions on a 0–10 scale, with scores below 7.0 marking a failure. JSONL format, organized by domain.

View on Hugging Face →

Get started: MCP server docs · Open Source hub · API & SDK reference

← Back to all announcements