Impraint

Compatibility

What is supported in beta, what is coming, and known gaps.

Support matrix

IntegrationStatus
Gemini on Vertex AI (LangChain, LangGraph, raw google-genai)Supported in beta
Gemini on AI Studio (LangChain, LangGraph, raw google-genai)Supported in beta
Anthropic (raw SDK)Coming after beta
OpenAIComing after beta
Claude on cloud gateways (Vertex AI, Bedrock, Azure)Coming after beta
OpenAI on AzureComing after beta
CrewAI, AutoGenPlanned
LlamaIndexNot yet planned

Features in beta

FeatureBeta scope
TracesGemini, LangChain and LangGraph
Content safetyPrompt checks (on_input, on by default in report mode) for Gemini, LangChain and LangGraph. Country packs: India and US. guard.check_content is available for any other code.
Rules and circuit breakerAny Python function, through @guard.tool or guard.intercept, whatever provider you use
ApprovalsDecide in the dashboard or in Slack. Slack buttons work through the connection Impraint sets up.
AlertsEmail and Slack. PagerDuty is not available.

Compliance reports

Beta includes two frameworks: DPDP and OWASP. Others, such as the EU AI Act, GDPR and SOC 2, are not part of beta.

Known gaps

  • Streaming for OpenAI has not been audited yet.
  • Reasoning and thinking-token accounting is verified for Gemini only.
  • Prompt checks on the Anthropic and OpenAI wrappers exist but are not tested or supported in beta.
  • Data retention: Guard does not delete data automatically in beta, even though Settings shows a retention period.

Tested versions

Guard is tested end to end (raw google-genai, LangChain and LangGraph; sync, async and streaming; Vertex AI and AI Studio) at these minimum versions:

PackageMinimum
langchain-core1.6.0
langgraph0.4.5
langchain-google-genai4.0.0
google-genai1.53.0

It also works on newer releases. Python 3.10 or newer is required.

Vertex AI with LangChain

Using Vertex AI through ChatGoogleGenerativeAI needs langchain-google-genai 4.0.0 or newer. Version 2.x supports AI Studio only, and 3.x ignores the Vertex settings, so neither is supported.

On this page