Key Concepts

Key Concepts

Understanding these core concepts will help you get the most out of Optima Forge. The platform is built on 29 architectural layers organized into 26 pillars, but you only need to understand a few key ideas to get started.

Providers

A provider is any LLM service that Forge can route to. Forge supports 14+ providers out of the box including OpenAI, Anthropic, Google Gemini, Mistral, Groq, Together, Fireworks, DeepSeek, Cohere, and more. Each provider is registered in the Provider Registry (Layer 1) with its models, capabilities, pricing, and rate limits. You can add your own API keys for each provider or use Forge's pooled keys.

Routing

Routing is how Forge decides which provider and model handles each request. The system uses a cascading intent classifier (regex/keyword, semantic-router, then LLM fallback) to understand query complexity, followed by a Quality Router that ranks models by ELO score, cost, latency, and capability. Setting model: "auto" enables full intelligent routing. You can also pin to a specific model like "gpt-4o" or "claude-sonnet-4-20250514".

Memory Layers

Forge provides three layers of persistent memory that work together:

• Vector Memory (Qdrant): Semantic search over past conversations and documents. Finds contextually relevant information using embeddings.
• Graph Memory (Neo4j + Graphiti): Relationship-aware knowledge graph that tracks entities, facts, and connections across conversations.
• State Memory (Redis CRDTs): Real-time session state using conflict-free replicated data types for cross-provider continuity.

Security Stages

ForgeGuard provides a 7-layer security pipeline (plus a Pre-Gate) that processes every request:

• Pre-Gate: SpiceDB + OPA authorization
• S1 — Gate: Rate limiting + Agent-Vault credential protection
• S2 — Input Scan: LlamaFirewall, DeBERTa-v3, LLM Guard, Presidio PII detection
• S3 — Vector Protection: Embedding-level threat detection
• S4 — Secure Route: Provider-aware security routing
• S5 — Output Scan: Response validation and PII scrubbing
• S6 — Audit: Augustus adversarial probing
• S7 — MCP Security: mcp-scan for tool call safety

Tiers

Forge offers four pricing tiers:

• Free: 100 requests/day, basic routing, limited memory
• Pro ($49/mo): 10,000 requests/day, full routing, memory, basic security
• Ultimate ($149/mo): Unlimited requests, all features, multi-agent, priority support
• Enterprise ($36K-$150K+/yr): Dedicated infrastructure, SSO/SCIM, SLA, custom compliance

Every new account receives a 7-day Ultimate trial that begins on your first API request.

MCP Modules

Forge exposes 55 MCP (Model Context Protocol) modules that can be activated a la carte. These allow AI agents to interact with Forge features programmatically. Modules are billed per-use via x402 micropayments or included in subscription tiers.

ForgeBot Agents

ForgeBots are custom AI agents built on the Forge platform using the ForgeBot Builder. Each bot can have multiple specialist agents coordinated by a Prime agent using the prime-first routing pattern. ForgeBots can connect to 3,000+ external APIs via Forge Connect, have persistent memory, and run security-scanned interactions.

Forge Connect

Forge Connect (Layer 29) provides universal API connectivity through a provider-agnostic abstraction layer. The default adapter uses Pipedream Connect, giving access to 3,000+ APIs and 10,000+ pre-built tools. Connected accounts and actions are secured through ForgeGuard and SpiceDB permissions.