Features

Hadrian Gateway includes a comprehensive feature set for production AI deployments. All features are free and included in the open-source release, dual-licensed under Apache 2.0 and MIT.

LLM Providers

Route requests to any major LLM provider through a unified OpenAI-compatible API.

Provider capabilities:

• Circuit breaker - Automatically disable unhealthy providers after repeated failures
• Automatic retry - Exponential backoff for transient errors (429, 5xx)
• Provider fallbacks - Chain providers: try A, then B, then C
• Model fallbacks - Graceful degradation: gpt-4o → gpt-4o-mini → claude-sonnet
• Health checks - Background monitoring to detect issues before user requests fail
• Model aliases - Create shortcuts like sonnet → claude-sonnet-4-20250514

Multi-Tenancy

Hadrian supports a flexible multi-tenancy hierarchy for organizations of any size.

Organizations
  └── Teams (optional)
        └── Projects
              └── Users
                    └── API Keys

Each level in the hierarchy can have:

Resource ownership:

Resources can be owned at different levels depending on your organization's needs:

• Organization-level: Shared across all teams and projects
• Team-level: Shared within a team
• Project-level: Isolated to a specific project
• User-level: Personal resources (conversations, API keys)

Authentication & Authorization

Flexible authentication supporting multiple methods for API and UI access.

CEL-based authorization:

Use Common Expression Language (CEL) policies for fine-grained access control:

[[auth.rbac.policies]]
name = "org-admin"
condition = "'admin' in subject.roles && context.org_id in subject.org_ids"
effect = "allow"

Guardrails

Block, warn, log, or redact content using configurable guardrails on both input and output.

Guardrail Providers

Execution Modes

Actions on violation:

• block - Reject request with error
• warn - Log warning, allow request
• log - Silent logging only
• redact - Remove/mask violating content

Budget Enforcement

Prevent cost overruns with atomic budget reservation and real-time enforcement.

How it works:

1. Request arrives → Reserve estimated cost ($0.10)
2. Forward to LLM provider
3. Request completes with actual cost
4. Adjust: Replace estimate with actual cost

This atomic reservation pattern prevents overspend even with concurrent requests.

Budget scopes:

• Organization-level budgets
• Team-level budgets
• Project-level budgets
• User-level budgets

Forecasting:

Built-in time-series forecasting with augurs:

• Projected spend for current period
• Days until budget exhaustion
• 95% confidence intervals

Vector Stores & RAG

OpenAI-compatible Vector Stores API for building RAG (Retrieval-Augmented Generation) applications.

Capabilities:

• Upload and process files (PDF, DOCX, TXT, Markdown, HTML, etc.)
• Automatic text extraction with OCR support via Kreuzberg
• Configurable chunking strategies (auto, fixed-size)
• Vector search with similarity scoring
• LLM-based re-ranking for improved relevance
• File search tool integration for Responses API

Vector backends:

Chat UI

Built-in React UI for multi-model conversations and administration.

Chat features:

• Multi-model comparison in single conversation
• Model instances (compare same model with different settings)
• Streaming markdown with syntax highlighting
• File uploads (images, PDFs)
• Conversation history with IndexedDB persistence
• Per-model settings (temperature, max tokens)

Chat modes:

Frontend Tools

Client-side tool execution in the browser via WebAssembly.

Tool results are displayed inline as interactive artifacts and sent back to the LLM to continue the conversation.

MCP Integration

Connect to external tool servers using the Model Context Protocol (MCP).

Capabilities:

• Connect to MCP servers via Streamable HTTP transport
• Automatic tool discovery from connected servers
• Tool execution with result streaming
• Persistent server connections across conversations

Use cases:

• File system access and manipulation
• Database queries
• External API integrations
• Custom enterprise tools

Response Caching

Cache LLM responses to reduce costs and latency.

Observability

Comprehensive monitoring and debugging capabilities.

[observability]
logging.format = "json"
logging.level = "info"

[observability.tracing]
enabled = true
exporter = "otlp"
endpoint = "http://localhost:4317"

[observability.metrics]
enabled = true

Data Privacy & GDPR

Built-in compliance features for data protection regulations.

Capabilities:

• Self-service data export (GDPR Article 15 - Right of Access)
• Self-service account deletion (GDPR Article 17 - Right to Erasure)
• Configurable data retention policies
• CSV export reports for compliance audits
• Audit logging for all privacy operations

Additional Features