LUMI

UNIFIED MULTI-PROVIDER GATEWAY Route requests to Anthropic Claude, OpenAI, AWS Bedrock, Google Vertex AI, Azure OpenAI, and Google GenAI through a single API. Full Claude API and OpenAI API compatibility - switch providers without code changes. Built-in health checks, automatic failover, and model discovery.

MCP GATEWAY Connect MCP (Model Context Protocol) servers once and expose their tools to all users and agents. Supports Global and Personal servers with secure ECDSA key exchange, real-time tool discovery via SSE, heartbeat monitoring, and encrypted credential storage. Agents integrate once and access all enabled tools.

INTELLIGENT CACHING Automatic prompt-level caching with adaptive boundary calculation eliminates redundant LLM calls. Cache profitability analysis applies caching only when cost-effective. Supports Bedrock-native cache TTLs. Tracks hit rates, miss costs, and generates per-user Cache Hero Scores to gamify optimization.

12-LAYER ANALYTICS ENGINE Basic: requests, tokens, spend, latency, errors, success rates. Personal: token ratios, latency experience, cache efficiency. Team: member performance, budget burn, model mix. Organization: cross-team ROI, vendor risk. Security: API key hygiene, jailbreak detection, anomalies. FinOps: context saturation, cache miss costs, spend forecasting. Agentic: conversation depth, tool chaining, agent failures. Prompt: reusability, compression, complexity. Efficiency: session productivity, hours saved. Session Quality: context-aware scoring. RAG/DX: error impact, rate limit forecasting. Real-Time: WebSocket live dashboards.

COST OPTIMIZER WITH AI RECOMMENDATIONS Simulate rate limits and budget caps against historical usage before applying them. AI-powered optimization suggestions via Claude. Preset configurations for Anthropic and OpenAI tier limits. Detailed breakdowns by model, user, and time period.

PROMPT EVALUATOR AND PROMPT MENTOR 6-dimension scoring: Clarity, Context, Structure, Output Specification, Reasoning, Safety - with letter grades A+ through F. AI-powered rewrite suggestions and best practice detection. Interactive Prompt Mentor for real-time guidance. Team and organization leaderboards.

SKILLS HUB - PLUGIN MARKETPLACE Create, version, publish, and distribute reusable AI skill plugins. Git-based distribution with YAML metadata. Public and private visibility, per-API-key activation, team sharing. AI-generated metadata. Direct Claude Code plugin integration. Build once, deploy across the organization.

6-LEVEL ROLE-BASED ACCESS CONTROL Permission hierarchy: Personal (L1), Team (L2-L3), Department Lead (L4), Org Admin (L5), Super Admin (L6+). Controls analytics access, model tiers, admin functions, spending visibility. Supervisor impersonation in read-only and write mode with full audit logging of every session and action.

ENTERPRISE SECURITY AES-256-GCM encryption for all credentials and API tokens at rest. Zero-knowledge admin access - admins see only masked tokens. Secure token separation: real tokens for auth, analytics tokens for logging. Per-user and per-key rate limiting with auto-unblock. CROWD SSO and Google OAuth.

TEAMS AND ORGANIZATIONS Role-based teams with owner, admin, member, and viewer roles. Share skills, budgets, and resources across teams. Department leads and org admins get aggregated views. Global View for super admins shows metrics across all users.

API KEY MANAGEMENT Named keys with per-key rate limits and budget caps. Key rotation preserves full analytics history. Admin key management via impersonation mode. Deactivation, reactivation, and lifecycle tracking.

FILE STORAGE AND SHARING Project-based file storage with sharing and access control. File overwrite preserves sharing permissions. Auto-registration of users from corporate directory. Direct browser access via web URLs.

SESSION AND CODE METRICS Automatic session management with duration tracking and quality scoring. Git diff integration tracks AI vs human lines per session. Jira task association for development productivity reporting.

REMOTE DESKTOP INTEGRATION WebSocket relay with Noise Protocol encryption (Noise_IK_25519_ChaChaPoly_BLAKE2s). Session sharing, collaboration, and remote UI through secure HTTP tunnels.

FEEDBACK AND RELEASE NOTES Kanban-based feedback workflow for suggestions, bugs, and improvements. AI-generated release notes from completed items using Claude Haiku. Soft-delete archival for audit trail.

AWS MARKETPLACE NATIVE Built-in subscription management, customer lifecycle tracking, metering, and billing event integration.