gensecaihq/pfsense-mcp-server

---

You:     Block all traffic from 203.0.113.5 on WAN
Claude:  ✓ created block rule  →  ✓ applied changes  →  rollback: restore_config_backup(revision_id=42)

You:     Why can't 192.168.1.50 reach the internet?
Claude:  ran diagnostics → gateway WAN_DHCP is down, and a block rule on LAN matches this host

You:     Add a WireGuard peer for my laptop and show me the config
Claude:  ✓ created peer on tun_wg0  →  here's the client config to import

pfSense MCP Server connects Claude Desktop, Claude Code, and any other MCP client to your pfSense firewall. Ask questions, diagnose issues, and change configuration through conversation — with a confirmation gate, config backup, and rollback on every destructive action.

Letting an AI touch a production firewall is only safe if the plumbing is right, so that’s where the work went: every tool’s wire format is verified against the pfSense REST API schema by a contract-test layer, and every change runs through a guardrail pipeline. 483 tests plus a wire-protocol E2E suite in CI on Python 3.11–3.13.

[!TIP] Jump to the Quick Start — about two minutes with uvx, no clone required. And if this saves you a trip through the pfSense web UI, a ⭐ helps others find it.

Contents

Why this exists · Quick start · What you can do · Safety · Supported versions · Authentication · Deployment · Configuration · Testing · MCP compliance · Architecture · Contributing

Why This Exists

Managing a pfSense firewall means clicking through web UI tabs, remembering field names, and hoping you don’t fat-finger a rule that locks you out. With this MCP server, you describe what you want in plain English and the AI handles the REST API calls, validates inputs, and warns you before anything destructive happens.

What makes it different:

• Every destructive operation requires explicit confirmation and shows you exactly what will happen
• Config backup before every delete/reboot — with a one-line rollback command (and an explicit warning if a backup point can’t be captured)
• Rate limiting on every mutating tool prevents runaway AI loops from flooding your firewall
• Positive input validation (IP/port/MAC/CIDR) plus path-traversal/XSS screening, and secrets redacted from logs and API error responses
• Wire-format verified against the pfSense REST API v2.10.0 schema by a contract-test layer, so tools send exactly what the API expects

Quick Start

Prerequisites: Python 3.11+, pfSense with REST API v2 package installed

Option A — run without cloning (uvx):

uvx --from git+https://github.com/gensecaihq/pfsense-mcp-server pfsense-mcp-server

Option B — clone for development:

git clone https://github.com/gensecaihq/pfsense-mcp-server.git
cd pfsense-mcp-server
pip install -r requirements.txt
cp .env.example .env
# Edit .env: set PFSENSE_URL, AUTH_METHOD, and credentials

Connect to Claude Desktop — add to ~/Library/Application Support/Claude/claude_desktop_config.json.

Using the installed entry point (Option A):

{
  "mcpServers": {
    "pfsense": {
      "command": "uvx",
      "args": ["--from", "git+https://github.com/gensecaihq/pfsense-mcp-server", "pfsense-mcp-server"],
      "env": {
        "PFSENSE_URL": "https://192.168.1.1",
        "AUTH_METHOD": "basic",
        "PFSENSE_USERNAME": "admin",
        "PFSENSE_PASSWORD": "your-password",
        "PFSENSE_VERSION": "CE_2_8_1",
        "VERIFY_SSL": "false"
      }
    }
  }
}

Or running from a clone (Option B):

{
  "mcpServers": {
    "pfsense": {
      "command": "python3.11",
      "args": ["-m", "src.main"],
      "cwd": "/path/to/pfsense-mcp-server",
      "env": {
        "PFSENSE_URL": "https://192.168.1.1",
        "AUTH_METHOD": "basic",
        "PFSENSE_USERNAME": "admin",
        "PFSENSE_PASSWORD": "your-password",
        "PFSENSE_VERSION": "CE_2_8_1",
        "VERIFY_SSL": "false"
      }
    }
  }
}

Start talking to your firewall. Open Claude Desktop and ask:

• “Show me all blocked traffic in the last hour”
• “What services are running?”
• “Create a port forward for port 443 to 192.168.1.50”
• “Run a full system health check”

What You Can Do

333 tools across every major pfSense subsystem:

Safety First

AI managing a production firewall needs guardrails. This server has 9 layers:

"Delete firewall rule 5"

  1. CLASSIFY    → HIGH risk (destructive)
  2. ALLOWLIST   → tool is permitted
  3. SANITIZE    → parameters clean (no injection)
  4. RATE LIMIT  → under 10 deletes/minute
  5. DRY RUN?    → user can preview first
  6. CONFIRM     → blocked until confirm=True
  7. BACKUP      → config revision captured
  8. EXECUTE     → API call made
  9. AUDIT LOG   → action recorded with redacted params

Response includes:
  "config_backup": {
    "pre_change_revision_id": 42,
    "rollback_instruction": "restore_config_backup(revision_id=42, confirm=True)"
  }

Every one of the 202 mutating tools carries a guardrail, enforced at registration by a meta-test so a new tool can’t ship ungated: the 52 destructive (delete/reboot/halt) tools require confirm=True, and the other 150 (create/update/apply/manage/export/service-control) are rate-limited, audited, and allowlist-checked. Sensitive parameters (passwords, keys, PSKs, bind passwords, tokens) are redacted in the audit log and in echoed API error responses.

You can also:

• Pass dry_run=True to preview any destructive operation without executing
• Pass verify_descr="Allow HTTPS" to verify you’re deleting the right rule (guards against ID shifts)
• Set MCP_READ_ONLY=true to expose only the 131 read-only tools (search, get, diagnose)
• Set MCP_ALLOWED_TOOLS=search_firewall_rules,get_firewall_log to restrict to specific tools

See SECURITY.md for the vulnerability-disclosure policy and deployment-hardening guidance.

Supported pfSense Versions

Requires the pfSense REST API v2 package by jaredhendrickson13. Package v2.8.x+ ships builds only for CE 2.8.1 and Plus 25.11.1/26.03/26.03.1; v2.7.3 is the last release with builds for CE 2.8.0 and Plus 24.11/25.11.

Security note: run REST API package v2.10.0+. It fixes a command-injection flaw in the interface-group endpoints (GHSA-w3w4-mvcc-vmgr) and adds core command auto-escaping; v2.9.0 fixed an earlier settings-sync privilege escalation (GHSA-8q8g-9f77-8g8g).

v2.10.0 also marks OpenVPNClient.auth_pass, User.ipsecpsk, and WireGuardPeer.presharedkey as sensitive, so the API no longer returns them by default. This server still sets them normally; if a workflow needs to read one back, add a sensitive-field override in the REST API settings.

Authentication

Three methods supported (configure in .env):

Deployment Options

stdio (default) — for Claude Desktop and Claude Code:

python3 -m src.main          # from a clone
pfsense-mcp-server           # via the installed console entry point (pip/uvx/pipx)

HTTP — for remote access and multi-client setups:

python3 -m src.main -t streamable-http --port 3000

Docker — hardened container with read-only filesystem:

docker compose up

Container security: non-root user (mcp:1000), read-only filesystem, all capabilities dropped, noexec tmpfs, no-new-privileges. In HTTP mode the container health check probes an unauthenticated /health endpoint (the /mcp endpoint requires a bearer token).

Configuration

Testing

python3 -m pytest tests/ -v          # 483 tests
python3 -m pytest tests/ --cov=src   # with coverage (~46%)

The suite includes a wire-contract layer (tests/contract/) that asserts every tool’s payload against the real pfSense REST API v2.10.0 schema (distilled from the upstream OpenAPI spec), so a wrong field name or type is a failing test rather than a silent misconfiguration. CI runs on Python 3.11/3.12/3.13 with pip-audit dependency scanning.

On top of the in-process suite, an end-to-end protocol smoke test drives the server over the real MCP wire protocol with the official MCP Inspector CLI — on both transports, in CI on every push:

make test-e2e            # or: ./scripts/inspector_smoke.sh  (needs node/npx, jq)

It verifies the initialize handshake, the 333-tool listing with annotations, the guardrail confirm-gate over the wire, read-only mode, and HTTP bearer-auth plus Origin enforcement — no pfSense instance required.

MCP Specification Compliance

Compliant with MCP 2025-11-25 — the newest revision with stable SDK support — and negotiates down to older revisions per connection, so existing clients keep working:

• ToolAnnotations on all 333 tools (readOnlyHint, destructiveHint, idempotentHint)
• serverInfo.version and instructions provided
• Origin header validation (MUST requirement)
• Bearer token auth with timing-safe comparison
• Default bind to localhost per spec SHOULD
• stdio and Streamable HTTP transports

The newest revision, MCP 2026-07-28, makes the protocol stateless; SDK support ships in fastmcp 4, currently in beta. This codebase already passes its full test suite on the fastmcp 4 beta (checked continuously by a non-blocking CI job), and uses none of the features 2026-07-28 deprecates (Roots, Sampling, MCP Logging), so adopting the sessionless protocol when fastmcp 4 is stable is a dependency-pin change. fastmcp 4 servers negotiate the protocol era per connection, keeping today’s handshake-era clients fully supported.

Project Structure

src/
  main.py              Entry point (transports, read-only filter, key validation)
  server.py            FastMCP instance + API client
  client.py            pfSense REST API v2 HTTP client (retry/backoff, pooling)
  guardrails.py        Risk classification, confirm gate, rate limit, audit, redaction
  helpers.py           Validation, parsing, pagination, safety guards
  models.py            Data models
  middleware.py        HTTP bearer auth + Origin validation + /health
  tools/               34 tool modules (333 tools)
scripts/
  generate_contract.py Regenerate the wire contract from an OpenAPI spec
  generate_token.py    Generate a secure MCP_API_KEY bearer token
  inspector_smoke.sh   End-to-end MCP protocol smoke test (MCP Inspector CLI)
tests/                 483 tests (incl. tests/contract/ wire-contract suite)

See ARCHITECTURE.md for the request lifecycle, guardrail model, and wire-contract layer; SECURITY.md for disclosure and hardening; and RELEASE_AUDIT.md for the audit and roadmap.

Contributing

We need real-world testing across diverse pfSense environments. See CONTRIBUTING or:

1. Fork and create a feature branch
2. Run python3 -m pytest tests/ -v
3. Submit a PR

Ideas: integration tests against real pfSense, additional package support (Snort, Suricata), Ollama local LLM bridge, multi-instance management.

License

MIT

Acknowledgments

• jaredhendrickson13 / pfrest — pfSense REST API v2 package
• JeremiahChurch — modular rewrite (PR #5), log endpoint OOM safeguards (PR #6)
• shawnpetersen — API v2 endpoint discovery (PR #3)
• aemitic — DELETE-body fix (PR #9), firewall ipprotocol for IPv6/dual-stack (PR #10), logconfigchanges (PR #11)
• hossamnagy — resilient startup on transient preflight failure (PR #14)
• bill-mccormick-dg — independent DELETE-body fix (PR #16)
• w1ld3r — DELETE and remote-syslog bug reports (#12, #13)
• tvlc — WebGUI port type-mismatch report (#7)
• renanwilliam — uvx/pipx packaging request (#8)
• Netgate — pfSense
• FastMCP — MCP framework