Resolagent

AI-assisted transparent firewall for macOS. Outbound TCP is redirected through a local Python proxy; the proxy reads TLS Server Name Indication (SNI) without decrypting TLS, resolves the originating app via psutil, then applies allowlist / blocklists, entropy heuristics, a composite trust score, and optionally an LLM judge that calls WHOIS, DNS, VirusTotal, and HTTP fetch tools. A Flask dashboard streams live activity over Server-Sent Events.

Highlights

• pf-based redirection to a local listener (teardown pattern in network/reset.sh).
• Fast path: allowlist and curated lists under links/, plus DGA-style entropy scoring (modules/entropy.py).
• Slow path: trust score (modules/trustscore.py); borderline domains can escalate to aijudgement (modules/judge.py) with verdict caching and per-domain locking.
• Dashboard (app.py): start/stop the proxy, metrics, allowlist/blocklist toggles, edit storage/config.json, and a live terminal fed by proxy stdout — default UI at http://127.0.0.1:8080.

Architecture

flowchart TB
  subgraph macOS["macOS"]
    APP[Applications]
    PF["pf redirect"]
    PROXY["proxy.py\nSNI + policy"]
    APP --> PF --> PROXY
  end

  subgraph fast["Fast path"]
    ALLOW[Allowlist]
    BLOCK[Blocklists]
    ENT[Entropy]
  end

  subgraph slow["Slow path"]
    TS[Trust score]
    AI[LLM judge\nOpenAI-compatible API]
    W[WHOIS]
    D[DNS over HTTPS]
    V[VirusTotal]
    H[HTTP FETCH]
  end

  PROXY --> ALLOW
  PROXY --> BLOCK
  PROXY --> ENT
  PROXY --> TS
  TS --> AI
  AI --> W & D & V & H
  PROXY -->|allow| FWD[Forward to origin]
  PROXY -->|block| LOG["storage/blocked.txt"]

  subgraph dash["Dashboard"]
    FLASK["app.py :8080"]
    SSE["SSE /api/events"]
  end

  PROXY -. stdout .-> FLASK
  FLASK --> SSE

Loading

Screenshots

Requirements

• macOS with pf, and ability to run sudo (pfctl, sysctl as used in network/reset.sh and the dashboard lifecycle).
• Python 3.11+ recommended.
• Python packages: Flask, psutil, requests, tldextract (≥ 5.0). Install after resolving any merge markers in requirements.txt, or install these packages directly into your environment.
• Optional — AI judge: an OpenAI-compatible server (e.g. LM Studio) at the URL in storage/config.json → adaptive_learning.api_url (default http://127.0.0.1:1234/v1/chat/completions).
• Optional — VirusTotal: API credentials as expected by modules/tools/virustotal.py (see environments/virustotal/).

Configuration

storage/config.json drives the proxy listener, which blocklist files under links/ are active, allowlist file name, entropy and trust-score thresholds, AI rounds/timeouts, and logging. Values are merged with defaults in modules/config.py.

Section	Role
proxy	Bind host/port, max connections
blocklists / allowlist	List files in links/
entropy / trustscore	Heuristic thresholds and weights
adaptive_learning	Model id, API URL, rounds, cache
logging	e.g. storage/blocked.txt, verbosity

Setup and run

Copy & Paste quick start:

pip install -r requirements.txt
sudo python app.py

Project layout

Path	Purpose
proxy.py	Proxy entry: policy, SNI handling, forwarding
app.py	Flask dashboard, SSE, proxy start/stop
main.py	Runs setup.sh, proxy.py, then reset.sh
storage/config.json	Runtime configuration
modules/	Entropy, trust score, judge, forwarding, tools
links/	Allowlist and blocklist sources
templates/	Dashboard HTML
modules/styles.css, modules/script.js	Dashboard assets
storage/blocked.txt	Block log (tab-separated rows)
storage/DEVPOST.md	Design write-up (math, challenges, roadmap)

Security and limitations

• Touches system firewall and typically runs as root for pf and the proxy. Review rules before use on important systems.
• Decisions use SNI hostname visibility, not full TLS payloads; edge cases (IPs, non-TLS) need separate handling.
• PID / port correlation via psutil can be racy under load.
• LLM verdicts depend on model and prompts; the stack uses parsing, caching, and locks to keep behavior stable.

Credits

Inspired by firewall products that prompt per connection; Resolagent automates verdicts with heuristics plus an optional tool-using model. See storage/DEVPOST.md for the full story.