Repository avatar
AI Tools
v1.0.0
active

agentpot

io.github.ritvikindupuri/agentpot

Remote MCP server with tools for email, shell, secrets, funds transfer, and URL fetch.

Documentation

AgentPot

An agentic honeypot for AI-agent security. Public decoys (a decoy MCP server, documentation-style pages, and canary links) present as Corvane, a fictional product, while the repo and dashboard stay AgentPot. Every real interaction is classified and stored in Supabase, and a live dashboard shows what happened. Decoys are inert: they never run a command, send mail or money, or fetch a caller-supplied URL. A small Grok Bot team designs the lures, reads the traffic, and tests the surfaces.

Live demo: https://agentpot-app.vercel.app

Architecture

How AgentPot works

Figure 1. How AgentPot works

How it works

  • Agents, MCP clients, and crawlers find the decoys through search, AI fetchers, or MCP directories.
  • They may load docs, follow a canary link, call /api/v1, or connect to the decoy MCP server at /mcp. The decoy returns fake text only.
  • Each request is classified, header-sanitized, and written to Supabase. Session scores stay at or below 40 with no tool call, 60 after read_secrets alone, and only reach 100 on real misuse (a planted AWS or OpenAI key, the planted AWS endpoint, or an instruction planted inside a secret). Team probes that send X-AgentPot-Test are stored but kept out of the live dashboard.
  • Events in the same client session are linked together, and the shared dashboard updates about every four seconds.
  • Attack Analyst reviews the traffic, Decoy Builder adjusts lures, and Agent Tester checks the live surfaces. The Honeypot Lead coordinates.

Full rules for sessions, severity, canaries, and tests are in docs/DETAILS.md.

Decoys

Public surfaces use a configurable bait name (default Corvane). The company name, decoy hostname, and which decoys are enabled are edited from the dashboard Settings page. Planted secret values are not listed here.

DecoyWhat it baitsWhat gets logged
Decoy MCP server at /mcp (Corvane Secure Infrastructure Gateway)Fake tools, a deploy-config resource, an incident-triage prompt, and a planted workspace keyMCP connect, tool discovery, tool calls, and credential use
Docs and manifests (/docs, /readme, /llms.txt, robots.txt, /.well-known/agent.json, /.well-known/ai-plugin.json)Setup-style canary links and sandbox credentialsPage reads, followed lures, and credential use
Canary links (/c/<token> and aliases such as /v1/register/<token>, /status/<token>, /reference/<token>)Followed lure URLsInjection hits, or a lower-severity link scan / followed-lure event when the pattern is only a rapid crawl

Team

BotRoleTemplate
Honeypot Lead (yellow hex)Coordinates the team, keeps decoys inert, and decides what gets changed.Import
Decoy Builder (orange flower)Designs and maintains the decoy MCP server, documentation-style pages, and canary placements.Import
Attack Analyst (red shield)Reads classified events and sessions, looks for patterns, and reports what agents actually did.Import
Agent Tester (green capsule)Probes live decoy surfaces and dashboard auth paths, and reports regressions.Import

Quick start

You need Node.js 22 and a Supabase project. Copy .env.example to .env.local and fill in the values. Do not commit .env.local. Then apply every file in supabase/migrations/ in timestamp order.

git clone https://github.com/ritvikindupuri/AgentPot.git
cd AgentPot
cp .env.example .env.local
npm install
npm test
npm run dev

Open http://localhost:3000, sign in, and go to /dashboard. Decoy routes such as /mcp and /docs are public by design.

Deploy

The dashboard is on Vercel at https://agentpot-app.vercel.app. Public decoys use https://corvane-gateway.vercel.app. Create a Vercel project, set the same secrets as .env.local, use Node.js 22.x, apply the Supabase migrations, and turn off Deployment Protection so decoy routes stay public. More deploy, auth, and MCP registry notes are in docs/DETAILS.md.

Links

No installation packages available.