AI OS Features

Chat with 24 LLM providers through a unified interface. Switch between OpenAI, Anthropic Claude, Google Gemini, Groq, and more without changing your workflow.

• ◆One interface for all major LLM providers
• ◆Provider failover and fallback chains
• ◆Streaming responses with real-time token display
• ◆Conversation history with search and export

Agents can use tools and call functions, with configurable approval gates that let you review and approve or reject tool calls before they execute.

• ◆Built-in tool system for code execution, file I/O, web access
• ◆Approval gates for sensitive operations
• ◆Policy-based automatic approval rules
• ◆Full audit log of all tool calls

A layered memory architecture with five tiers: episodic (turn summaries with auto-decay), semantic (injected facts with hybrid FTS5 + vector search), reflection (LLM-extracted behavior patterns), graph (knowledge graph entities and relationships), and skills (procedural knowledge from sessions).

• ◆Episodic: Turn summaries with configurable half-life and auto-decay scoring
• ◆Semantic: Injected facts with hybrid FTS5 + cosine vector similarity search
• ◆Reflection: LLM-extracted behavioral patterns with weekly consolidation
• ◆Graph: Knowledge graph nodes and edges built from semantic entity extraction
• ◆Skills: Procedural knowledge extracted from sessions with multi-step tool usage

A multi-layered security architecture with encrypted vault storage, granular policy engine, role-based access control, and approval workflows.

• ◆Encrypted vault for secrets and credentials
• ◆Policy engine with allow/deny rules
• ◆Approval workflows for sensitive actions
• ◆Sandboxed plugin execution isolation

Execute Python, JavaScript, Wasm, and shell commands in isolated sandboxes. Perfect for data analysis, automation, and prototyping.

• ◆Python sandbox with pip package support
• ◆WebAssembly runtime with WASI support
• ◆Resource limits (CPU, memory, timeouts)
• ◆Read-only mode for untrusted scripts

Two intelligent routing strategies wrapping LLM providers. Cascade tries cheapest providers first and escalates on low confidence. Threshold scores prompt complexity to route between strong and weak models.

• ◆Cascade router: cheapest model first, escalate on low confidence
• ◆Threshold router: RouteLLM-style complexity scoring for prompt routing
• ◆Cost optimization with confidence-based fallback between providers
• ◆Graceful failover on provider outages

Persistent daemon mode with a built-in job scheduler. Run agents in the background, schedule recurring tasks, and monitor execution.

• ◆Persistent background daemon process
• ◆CRON-like job scheduling
• ◆Job queue with priority levels
• ◆Execution logs and monitoring

Extend CortexPrism with a powerful plugin system supporting ESM modules, Model Context Protocol servers, and WebAssembly plugins.

• ◆Three plugin types: ESM, MCP, WASM
• ◆Plugin marketplace for discovery
• ◆Version management with dependency resolution
• ◆Isolated plugin sandboxes

Built-in web interface and comprehensive REST API for managing agents, monitoring sessions, and configuring the system remotely.

• ◆Responsive dark-theme web dashboard
• ◆REST API with OpenAPI documentation
• ◆WebSocket for real-time streaming
• ◆Session management and monitoring

Create, configure, and manage multiple specialized agents with different personalities, tools, and provider configurations.

• ◆Multiple agent profiles with custom configurations
• ◆Agent-specific system prompts and souls
• ◆Per-agent tool and provider assignments
• ◆Agent-to-agent collaboration

Run CortexPrism as a distributed system with micro-service architecture. Scale components independently and deploy across multiple machines.

• ◆Distributed agent execution
• ◆Message queue integration
• ◆Service discovery and health checks
• ◆Horizontal scaling of components

Full git porcelain interface powered by the Cortex agent. Auto-generated commit messages, diff analysis, branch management, and PR-ready workflows.

• ◆Ten git subcommands: status, log, diff, add, commit, push, pull, clone, branch, remote
• ◆Agent-generated conventional commit messages from diffs
• ◆Diff analysis with natural language summaries
• ◆Intelligent branch naming and conflict detection

An adaptive 6-signal model-selection engine that observes LLM calls, learns patterns, and predicts the optimal model for each task. Powered by reinforcement learning.

• ◆Six prediction signals: trajectory, episodic, historical, cost, quality, reflection
• ◆Automatic tool execution above 90% confidence for safe operations
• ◆Reinforcement learning with EMA weight adjustment (alpha=0.15 reward, alpha=0.25 punishment)
• ◆Dashboard with accuracy bars, signal weights, and top tools

A DSL-based workflow engine for defining and executing multi-step agent tasks. Supports sequential steps, conditional branching, parallel execution, and human approval gates.

• ◆Five node types: step, branch, parallel, goto, wait
• ◆Conditional branching with if/then/else logic
• ◆Parallel execution via Promise.allSettled
• ◆Human-in-the-loop approval gates with resume support

A comprehensive audit and telemetry event log that records every action in the system — LLM calls, tool execution, security decisions, memory operations, and quartermaster predictions.

• ◆50+ event types tracked with session, actor, and time indexing
• ◆Immutable event log stored in dedicated lens.db SQLite database
• ◆Web UI activity dashboard with timeline view
• ◆REST API for programmatic audit log access

Auto-extracts reusable procedural patterns from successful tool-call sequences and stores them as versioned, quality-scored skills with trust tiering and dependency tracking.

• ◆6-state lifecycle: candidate → verified → released → degraded → deprecated → archived
• ◆Embedding-based retrieval with cosine similarity matching to user queries
• ◆Automatic deduplication and merging of similar skills
• ◆4-tier trust system controlling agent skill exposure

Tree-sitter WASM parser for 40+ languages with call graph resolution, impact analysis, architecture extraction, and symbol search. Visualized via D3.js force-directed graph.

• ◆40+ languages: TS, JS, Python, Go, Rust, Java, Kotlin, C, C++, C#, Ruby, PHP, Swift, Scala, Lua, Bash, SQL, and more
• ◆7-strategy call target resolution with cross-file import analysis
• ◆12 node labels and 18 edge types in the graph database
• ◆D3.js Web UI with symbol search, impact analysis, and path tracing

Full voice interaction pipeline with OpenAI Whisper STT, OpenAI TTS / ElevenLabs, energy-based VAD, and real-time audio streaming over WebSocket.

• ◆Speech-to-text via OpenAI Whisper with configurable models
• ◆Text-to-speech with OpenAI TTS and ElevenLabs provider support
• ◆Energy-based voice activity detection with adjustable threshold
• ◆Real-time audio streaming over WebSocket for low-latency interaction

GUI automation using X11 virtual framebuffer (Xvfb) with mouse control, keyboard input, and screenshot capture. Docker-isolated or native runtime.

• ◆Virtual display management with multi-display support
• ◆Mouse: coordinate movement, left/right/middle clicks, double/triple clicks, drag
• ◆Keyboard: text typing with configurable delays, key combos, key holding
• ◆15 agent actions: screenshot, click, type, key, scroll, wait, mouse_move, drag

Headless Playwright-powered browser for web navigation, interaction, screenshots, and accessibility snapshots. Fully integrated with agent tools.

• ◆Navigate to URLs, click elements, type text, and submit forms
• ◆Take full-page screenshots and accessibility snapshots
• ◆Execute JavaScript on pages and wait for conditions
• ◆All actions gated through policy validator with user approval

Google Agent2Agent (A2A) v1.0 protocol for cross-framework agent collaboration. JSON-RPC 2.0 server/client with SSE streaming and tool wrapping.

• ◆Full A2A v1.0 compliance with JSON-RPC 2.0 transport
• ◆Server mode: expose CortexPrism agents to other A2A-compatible frameworks
• ◆Client mode: consume external A2A agents as CortexPrism tools
• ◆SSE streaming for real-time agent-to-agent communication