SCOUT — Autonomous Vulnerability Research Agent

Local Ollama-based agent wired into VS Code AI extensions and a custom MCP server for smart contract + web2 security research.

Quick start

# 1. start the OpenAI-compat server (background)
python3 ~/scout/server.py --bg

# 2. point your VS Code AI extension at it
#    Continue:  http://127.0.0.1:11435/v1
#    Cline/Roo: http://127.0.0.1:11435/v1, model "scout-mem"
#    API key:   any string (e.g. "scout")

# 3. or use the CLI directly
python3 ~/scout/bridge.py "audit 0xABC... on Ethereum"

# 4. or run a full Polymarket audit pipeline
python3 ~/scout/polymarket_orchestrator.py 0xABC...

# 5. or use the MCP server (when an MCP client is wired in)
echo '{"jsonrpc":"2.0","id":1,"method":"tools/list"}' | python3 ~/scout/mcp_server.py

Files

File	Purpose
SYSTEM_PROMPT.md	The reference system prompt (LM Studio / manual)
VSCODE_SYSTEM_PROMPT.md	Same prompt, paste-friendly for VS Code extensions
Modelfile	Builds scout Ollama model (qwen2.5-coder:14b base)
Modelfile.mem	Builds scout-mem (all 3 base models + memory schema)
bridge.py	Tool implementations (file, shell, memory) — REPL entry point
server.py	OpenAI-compat HTTP server — VS Code extensions point here
mcp_server.py	Model Context Protocol server — exposes 8 security tools
mcp-manifest.json	MCP server manifest for registry publication
polymarket_orchestrator.py	5-phase audit pipeline for Polymarket targets
polymarket/SYSTEM_PROMPT.md	Polymarket-specific agent prompt
polymarket/AGENT_BRIEFINGS.md	6 sub-agent role briefings
polymarket/Modelfile.pm	Builds scout-pm (Polymarket specialized)
RESEARCH_2026-06.md	Verified tech landscape as of June 2026
memory/	Persistent state (user, environment, session JSON)

Ollama models

Model	Base	Size	Use
scout	qwen2.5-coder:14b	9 GB	Code generation, structured output
scout-mem	qwen2.5-coder + deepseek-r1 + gemma4	9.6 GB	Memory + tools, multi-task
scout-pm	qwen2.5-coder + Polymarket prompt	9 GB	Polymarket-specific audits

Switch the server's model:

kill $(cat ~/scout/server.pid)
SCOUT_MODEL=scout-pm python3 ~/scout/server.py --bg

MCP server

8 tools exposed:

• scout_recon — subdomain enum + contract discovery
• scout_scan_solidity — slither + aderyn
• scout_fuzz — echidna or medusa
• scout_cast_call — read-only on-chain calls (refuses state-changing methods)
• scout_decompile — heimdall-rs bytecode decompilation
• scout_search_findings — local finding store search
• scout_triage — Polymarket severity scoring
• scout_draft_report — markdown report generation

Wire into Continue (~/.continue/config.json):

{
  "mcpServers": [{
    "name": "scout",
    "command": "python3",
    "args": ["~/scout/mcp_server.py"]
  }]
}

Toolchain dependencies

Required for the security tools to work:

brew install foundry    # forge, cast, anvil
pipx install slither-analyzer
pipx install aderyn
# Optional:
pipx install eth-security-toolbox

Polymarket audit pipeline

# Full automated pipeline
python3 ~/scout/polymarket_orchestrator.py 0x4d97fc1d4d8b8b9b48f9e5d6c2a1b3f4e5d6c7a8b

# Or just the orchestrator with a web2 target
python3 ~/scout/polymarket_orchestrator.py --target https://polymarket.com

# Resume an interrupted session
python3 ~/scout/polymarket_orchestrator.py --resume 1734567890-a1b2c3

# Start from a specific phase
python3 ~/scout/polymarket_orchestrator.py --resume <id> --phase 4

Output goes to /tmp/scout/sessions/<id>/.

State persistence

Three memory files in ~/scout/memory/:

• user.json — operator profile (handle, platforms, preferences)
• environment.json — local setup (tools, paths, models)
• session.json — active target, scope, findings_so_far

Reset:

rm ~/scout/memory/*.json

Server management

# start
python3 ~/scout/server.py --bg

# status
curl http://127.0.0.1:11435/health

# stop
kill $(cat ~/scout/server.pid)

# logs
tail -f ~/scout/server.log

Known limitations

• 14B-class models are not reliable for novel exploit reasoning; for Critical-class findings, always verify manually
• Local models don't have web access; web2 recon is constrained to local tooling
• No memory across server restarts (until you commit to writing to memory)
• No parallel sub-agent execution; sequential tool calls per turn
• MCP server doesn't auto-publish to registry yet (you have to push mcp-manifest.json)

Roadmap

• Per-agent role-specific Ollama models (scout-coder, scout-reasoner, scout-comm)
• Memory bridge to Hermes Agent's memory schema
• Auto-publish to MCP registry
• Real parallel sub-agent fan-out via asyncio
• Heuristic-driven static analysis (custom detectors)