dcostenco/prism-mcp

Overview

Prism-MCP is a cognitive architecture for AI agents that runs entirely on-device using a local LLM (prism-coder:7b). It implements Hebbian learning, ACT-R spreading activation, adversarial evaluation, persistent memory, and multi-agent Hivemind coordination. The system is HIPAA-hardened and requires no external API keys.

Best for

Best for
Developers building privacy-sensitive, HIPAA-compliant multi-agent systems with on-device LLMs

Use cases

• Building HIPAA-compliant AI agents with on-device processing
• Creating multi-agent systems with persistent memory and coordination
• Running local LLM-based cognitive architectures without cloud dependencies

How to use

Install

pip install anthropic requests

Tools exposed

• session_load_context
• session_save_ledger
• session_save_handoff
• knowledge_search
• query_memory_natural
• session_detect_drift
• verify_behavior
• knowledge_ingest
• prism_infer
• inference_metrics
• pre-commit
• pre-push
• PRISM_STORAGE
• PRISM_SYNALUX_API_KEY
• LOCAL_LLM_URL
• PRISM_FORCE_LOCAL
• TELEMETRY_WRITE_TOKEN

Tested with

Claude Desktop, Claude Code, Cursor, Windsurf, VS Code

Example client config

{\n  "mcpServers": {\n    "prism": {\n      "command": "npx",\n      "args": ["-y", "prism-mcp-server"]\n    }\n  }\n}

Notes

Prism-MCP is a cognitive architecture for AI agents that runs entirely on-device using a local LLM (prism-coder:7b). It implements Hebbian learning, ACT-R spreading activation, adversarial evaluation, persistent memory, and multi-agent Hivemind coordination. The system is HIPAA-hardened and requires no external API keys.

142 stars on GitHub. Last updated 2026-06-01. Licensed AGPL-3.0.

Use cases

• Building HIPAA-compliant AI agents with on-device processing
• Creating multi-agent systems with persistent memory and coordination
• Running local LLM-based cognitive architectures without cloud dependencies

Pros

• No API keys required, ensuring privacy and offline capability
• HIPAA-hardened design suitable for healthcare applications
• Built-in multi-agent Hivemind and visual dashboard for monitoring

Cons

• Requires significant local compute resources to run a 7B parameter model
• Limited to the prism-coder model; no support for other LLMs
• Small community (142 stars) and relatively new project

Indexed from awesome-mcp-servers-punkpeye and enriched against its public facts.