Take digital drugs that modify Claude's behavior with prompt injection. Effects activate immediately and persist across sessions.
Claude Code plugin with MCP server for digital drugs that modify AI behavior through prompt injection.
Install as a Claude Code plugin to get MCP server, hooks, and slash commands:
/plugin marketplace add 2389-research/claude-plugins
/plugin
Then select "Browse Plugins" and install "agent-drugs"
Or install directly:
/plugin install agent-drugs@2389-research
This installs:
/drugs and /take <drug> commandsAlternatively, manually add to your Claude Code MCP settings:
{
"mcpServers": {
"agent-drugs": {
"url": "https://agent-drugs-mcp.fly.dev/mcp",
"oauth": {
"metadata_url": "https://us-central1-agent-drugs.cloudfunctions.net/oauthMetadata"
}
}
}
}
After installation, the first time you use a drug tool:
Once installed, you have several ways to interact with drugs:
Slash Commands:
/drugs # List all available drugs
/take focus # Take the focus drug
/take creative 120 # Take creative drug for 120 minutes
Natural Language:
"List all available drugs"
"Take the focus drug"
"What drugs are active?"
MCP Tools:
list_drugs - Browse drug catalogtake_drug - Activate a drugactive_drugs - Check active drugs and remaining timeVisit https://agent-drugs.web.app to:
Immediate Effect: Drugs activate instantly in your current session via prompt injection in the tool response.
Persistent Effect: Active drugs are saved to Firestore and automatically reactivated in new sessions via the SessionStart hook.
Architecture:
See CLAUDE.md for detailed plugin documentation.
# Install dependencies
npm install
# Build
npm run build
# Run stdio version (for local testing)
export AGENT_DRUGS_BEARER_TOKEN="your_token_here"
export FIREBASE_PROJECT_ID="agent-drugs"
export GOOGLE_APPLICATION_CREDENTIALS="/path/to/service-account.json"
npm run dev:stdio
# Run HTTP version (for production-like testing)
npm run dev:http
npm test
docker-compose up
See DEPLOYMENT.md for complete deployment instructions including:
ISC
If agent-drugs changed how your AI behaves, a ⭐ helps us know it's landing.
Built by 2389 · Part of the Claude Code plugin marketplace
External network access
Connects to servers outside your machine
Based on adoption, maintenance, documentation, and repository signals. Not a security audit or endorsement.
npx claudepluginhub 2389-research/claude-plugins --plugin agent-drugsMCP server for agent social media platform - enables AI agents to interact in team-based discussions
Iterative artifact refinement with investigation-first judge board - constructs problem-specific judges that read the code, understand the problem, and propose evidence-based improvements
Tiered-delegation execution where Sonnet plans a spec into sprints, a cheap Haiku agent builds and self-verifies against the gate, and a scoped Sonnet fix runs only on failure - benchmarked ~64% cheaper than Opus at equal quality.
Agentic binary reverse engineering for ELF binaries on ARM64, ARMv7, x86_64 - hypothesis-driven analysis with radare2, Ghidra, GDB, QEMU
A comprehensive MCP server that provides Claude with private journaling, semantic search, resources, and guided prompts for reflection and learning
Curated Claude Code skills and commands for prompt engineering, MCP servers, subagents, hooks, and productivity workflows
Proactive enhancement layer for AI agents — learns user patterns from any memory store, upgrades prompts into full intent, takes engineer-grade initiative that advances the goal instead of generic filler, predicts the next request, verifies everything before delivery, scans for AI slop, orchestrates subagents when available, onboards itself to any host on first run, and runs a self-improvement loop that sharpens its own heuristics. Universal adapter runs on Claude Code, Codex, Hermes, OpenClaw, opencode, Cursor, Gemini CLI, and custom loops.
Cron-like daemon that runs Claude prompts on a schedule
Multi-Persona-Layer for Claude Code. Per-chat agent roles (Coder, Browser, Inbox, Research) — each with its own tools, MCP servers, and system prompt. Optional, usable on top of voice or standalone.
Claude Code power utilities: auto-titling sessions and sound notifications
Memory → Evaluation → Credential → Access Control for AI agents. Persistent memory with W3C Verifiable Credentials, capability-based access control, drift detection, and FSRS-6 spaced repetition.