Frequently Asked Questions

General

What is CortexPrism?

CortexPrism is an open-source agentic harness system written in TypeScript/Deno. It provides a runtime for building, deploying, and managing AI agents with support for 12+ LLM providers, a 5-tier memory system, sandboxed code execution, and a defense-in-depth security model.

What is an "agentic harness"?

An agentic harness is a runtime system that hosts, orchestrates, and empowers AI agents. It manages the full lifecycle of agent interactions including LLM calls, tool execution, memory persistence, security validation, and user interfaces.

Do I need a GPU?

No. CortexPrism runs entirely on CPU. All LLM calls are made through provider APIs — you only need a network connection.

Is it free and open source?

Yes. CortexPrism is released under the MIT License. The source code is available on GitHub.

What are the system requirements?

  • OS: Linux, macOS, or Windows (via WSL2)
  • Runtime: Deno v2.0+
  • Memory: 4GB RAM minimum (8GB+ recommended)
  • Storage: 500MB for base installation
  • Optional: Docker for sandboxed code execution

LLM Providers

What providers are supported?

12+ providers are supported: Anthropic Claude, OpenAI, Google Gemini, Mistral AI, Groq, DeepSeek, OpenRouter, xAI (Grok), Together AI, AWS Bedrock, Cohere, and Ollama (local).

Can I switch providers mid-session?

Yes. Use the /model slash command inside a chat session to switch models, or start a new session with --model or --provider flags.

Can I use multiple providers at once?

Yes. The CascadeRouter can be configured to try cheaper models first and escalate to more capable ones when confidence is low.

Can I use CortexPrism offline?

CortexPrism requires network access for LLM API calls. However, you can run local models via Ollama for fully offline operation. The memory system, plugin management, and tool execution work offline.

Memory

How does the memory system work?

CortexPrism uses a 5-tier memory system with hybrid retrieval combining FTS5 keyword search and cosine vector similarity, scored with exponential time decay. Memories are automatically retrieved and injected into the agent's context.

Is my data private?

Memory data is stored locally in SQLite databases on your machine. The Vault uses AES-256-GCM encryption for sensitive credentials. No memory data is sent to external services except the LLM provider API calls.

Security

How does the Parallax security model work?

Every tool call passes through a 3-stage validator: tool name check, shell command pattern check, and domain allow/deny rules. All decisions are logged to the Cortex Lens audit trail.

What dangerous commands are blocked by default?

  • rm -rf / and variants
  • Shell fork bomb patterns
  • Direct disk writes (dd to block devices)
  • World-writable root permissions (chmod 777 /)