A Model Context Protocol (MCP) server that provides real-time access to Firewalla firewall data through 28 read-only tools, plus 11 opt-in write tools, compatible with any MCP client.
This is a community project, not affiliated with Firewalla. It reads your data through the Firewalla MSP API, so it needs an MSP account with API access.
Model Context Protocol (MCP) server for Firewalla MSP API - Provides real-time network monitoring, security analysis, and firewall management through 28 read-only tools, plus 11 opt-in write tools, compatible with any MCP client
2026-10-28
星數
—
工具
—
可用率
—
Why Firewalla MCP Server?
Simple Network Security Integration
28 read-only tools for network monitoring and analysis: 23 Direct API Endpoints + 5 Convenience Wrappers
11 write tools, off unless FIREWALLA_ENABLE_WRITE_TOOLS=true, so by default nothing can change your box
Advanced Search with query syntax and filters
Clean, Verified Architecture with corrected API schemas
Features
Real-time Firewall Data: Query security alerts, network flows, and device status
Security Analysis: Get insights on threats, blocked attacks, and network anomalies
Bandwidth Monitoring: Track top bandwidth consumers and usage patterns
Rule Management: View firewall rules; with the write tools, create, pause, resume and delete them
Target Lists: View target lists; with the write tools, create, update and delete them
Search Tools: Query syntax with filters and logical operators
The MCP server acts as a bridge between Claude and your Firewalla firewall, translating Claude's requests into Firewalla API calls and returning the results in a format Claude can understand.
Prerequisites
Node.js 18+ and npm. On Node 18-22, npm prints an EBADENGINE warning for geoip-lite 2.x, which declares Node 24 for its database update script; the lookups the server uses run on Node 18 and later, and CI tests 18, 20, 22 and 24.
Firewalla MSP account with API access
Your Firewalla device online and connected
Upgrading to 2.0.0
2.0.0 changes these defaults from 1.5.0. A setup that relied on an old default needs the setting in the last column.
Up to 1.5.0
From 2.0.0
What to set
pause_rule, resume_rule, create_target_list, update_target_list and delete_target_list were always registered
All 11 write tools are off; the server lists 24 read-only tools, and a call to a write tool answers "Unknown tool" and sends nothing
FIREWALLA_ENABLE_WRITE_TOOLS=true
The HTTP transport listened on every interface
It listens on 127.0.0.1. The Docker image sets MCP_HTTP_HOST=0.0.0.0, so a published port reaches it once MCP_HTTP_BEARER_TOKEN is set
MCP_HTTP_HOST=0.0.0.0 outside Docker, with MCP_HTTP_BEARER_TOKEN
Any Host header was served
403 unless it is localhost, 127.0.0.1, [::1] or the MCP_HTTP_HOST address; a wildcard MCP_HTTP_HOST (0.0.0.0, ::, as in the Docker image) adds nothing, so every other name clients use must be listed
the name clients connect by (a compose service name, a LAN address) in MCP_HTTP_ALLOWED_HOSTS
Any Origin was served
403 for a request with an Origin header, which browsers send; MCP clients that send none are not affected
the page's origin in MCP_HTTP_ALLOWED_ORIGINS: scheme, host and port (http://localhost:6274). The setting and the header are compared as URL origins, so letter case and a default port make no difference and a path in the setting is dropped. There is no wildcard: an entry with * in its host stops the server at startup. The opaque origin null is always refused
No token check (MCP_HTTP_BEARER_TOKEN did not exist)
On an address other than loopback, the Docker image's 0.0.0.0 included, the server exits with code 1 at startup unless MCP_HTTP_BEARER_TOKEN is set, and it refuses a token shorter than 16 characters wherever it listens. With the token, a request without Authorization: Bearer <token> gets 401
MCP_HTTP_BEARER_TOKEN (openssl rand -hex 32 makes one) and the same token in every client, or MCP_HTTP_ALLOW_NO_TOKEN=true on a network no untrusted machine can reach, such as a compose network with no published port
Any path starting with MCP_HTTP_PATH was served
Only the configured MCP_HTTP_PATH (default /mcp) and that path with one trailing slash, with or without a query string; with the default, /mcpx and /mcp/x get 404
a client URL that ends in exactly the configured MCP_HTTP_PATH
MCP_TEST_MODE=true started under any NODE_ENV
Refused with NODE_ENV=production, which the Docker image sets
-e NODE_ENV=development with -e MCP_TEST_MODE=true
API_RATE_LIMIT was range-checked and not applied
Each server process starts at most API_RATE_LIMIT requests in any rolling 5 minutes (default 100), counting 429 retries and the retry of a failed read; cache hits do not count. A request that cannot start within 20 s fails and says when capacity returns. A request still waiting when its tool gives up at 30 s is not sent; archive_alarm, mute_alarm and delete_alarm have no such tool deadline. The MSP API allowed 100 requests per fixed 5-minute window on the one token measured; whether that is per token or per account was not measured
when several server processes or other clients use one token, an API_RATE_LIMIT for each so that together they stay at 100 or less: each process counts only its own requests, so lowering one does not cover the others
Other changes a script may notice:
With FIREWALLA_BOX_ID set, get_alarm_trends covers that box instead of every box, and makes 31 requests for the default 30d instead of 1. An explicit group still reads the group's series in one request.
Tool responses are compact JSON, without the indentation.
AND, OR and NOT are translated to the API's grammar, which has none of them: status:blocked AND region:US matched 0 flows up to 1.5.0 and now matches what status:blocked region:US does. A query with no API form is refused as a validation error that suggests runnable queries: an OR between different fields, NOT over an AND, a comma list with a space around a comma, and geographic names the API does not have, such as country: (use region:). Up to 1.5.0 these returned wrong results or none.
IDs are no longer trimmed: an ID with leading or trailing whitespace, /, ?, # or % is refused before any request.
A streamed get_flow_data chunk gives each flow's time as ts, an ISO string, like a plain page, instead of timestamp.
Over HTTP, a request with the session ID of a session the server no longer holds gets 404 instead of 400.
# Install globally
npm install -g firewalla-mcp-server
# Or install locally in your project
npm install firewalla-mcp-server
Option B: Use Docker
Warning: Not for production use – secrets visible in process list
The examples below pass credentials directly in the command line, which exposes them to process listing and shell history. For production use, consider these secure alternatives:
Use --env-file with a .env file: docker run --env-file .env ...
Set environment variables in your shell before running Docker
Use Docker secrets for orchestration environments
Stdio Transport (Default - for Claude Desktop integration):
# Using Docker Hub image (minimal config)
docker run -it --rm \
-e FIREWALLA_MSP_TOKEN=your_token \
-e FIREWALLA_MSP_ID=yourdomain.firewalla.net \
amittell/firewalla-mcp-server
# Or with optional box filter
docker run -it --rm \
-e FIREWALLA_MSP_TOKEN=your_token \
-e FIREWALLA_MSP_ID=yourdomain.firewalla.net \
-e FIREWALLA_BOX_ID=your_box_gid \
amittell/firewalla-mcp-server
# Or build locally
docker build -t firewalla-mcp-server .
docker run -it --rm \
-e FIREWALLA_MSP_TOKEN=your_token \
-e FIREWALLA_MSP_ID=yourdomain.firewalla.net \
firewalla-mcp-server
# Recommended: Using env file (more secure)
docker run -it --rm --env-file .env amittell/firewalla-mcp-server
HTTP Transport (for standalone Docker containers and external access):
# Make a token and keep the value echo prints: every client sends it
export MCP_HTTP_BEARER_TOKEN="$(openssl rand -hex 32)"
echo "$MCP_HTTP_BEARER_TOKEN"
# Run with HTTP transport on port 3000. -e MCP_HTTP_BEARER_TOKEN with no
# value passes the exported one without putting it on the command line
docker run -d --name firewalla-mcp \
-p 3000:3000 \
-e MCP_TRANSPORT=http \
-e MCP_HTTP_PORT=3000 \
-e MCP_HTTP_BEARER_TOKEN \
-e FIREWALLA_MSP_TOKEN=your_token \
-e FIREWALLA_MSP_ID=yourdomain.firewalla.net \
amittell/firewalla-mcp-server
# Add FIREWALLA_BOX_ID if you want to filter to a specific box
# -e FIREWALLA_BOX_ID=your_box_gid \
# The server will be accessible at http://localhost:3000/mcp, and clients
# send the header: Authorization: Bearer <the token echo printed>
# Using env file (recommended): besides the credentials, .env sets
# MCP_TRANSPORT=http and MCP_HTTP_BEARER_TOKEN; to add a new token to it:
# echo "MCP_HTTP_BEARER_TOKEN=$(openssl rand -hex 32)" >> .env
docker run -d --name firewalla-mcp \
-p 3000:3000 \
--env-file .env \
amittell/firewalla-mcp-server
# For docker-compose
cat > docker-compose.yml << EOF
version: '3.8'
services:
firewalla-mcp:
image: amittell/firewalla-mcp-server
ports:
- "3000:3000"
environment:
- MCP_TRANSPORT=http
- MCP_HTTP_PORT=3000
# The image already listens on every interface of the container
- MCP_HTTP_HOST=0.0.0.0
# From the shell or a .env next to this file; openssl rand -hex 32
- MCP_HTTP_BEARER_TOKEN=\${MCP_HTTP_BEARER_TOKEN}
# Other containers reach it as http://firewalla-mcp:3000/mcp
- MCP_HTTP_ALLOWED_HOSTS=firewalla-mcp
- FIREWALLA_MSP_TOKEN=\${FIREWALLA_MSP_TOKEN}
- FIREWALLA_MSP_ID=\${FIREWALLA_MSP_ID}
# Optional: filter to specific box
# - FIREWALLA_BOX_ID=\${FIREWALLA_BOX_ID}
restart: unless-stopped
EOF
docker-compose up -d
The image sets MCP_HTTP_HOST=0.0.0.0 so that a published port reaches the server, and on that address the server does not start without MCP_HTTP_BEARER_TOKEN (at least 16 characters; openssl rand -hex 32 makes one): the container exits with code 1, and docker logs shows a refusing to start line that says why. -p 3000:3000 publishes the port on every interface of the Docker host, where anyone on your network can reach it. Publishing it on this machine only (-p 127.0.0.1:3000:3000) still needs the token: another container on the same Docker network can connect to the container's own address, and one that did, with Host: localhost and no token, got 200 when the check was turned off. MCP_HTTP_ALLOW_NO_TOKEN=true turns the check off, for a network no untrusted machine or container can reach, such as a compose network with no published port; the server then prints a warning when it starts. The server answers only requests whose Host header is localhost, 127.0.0.1 or [::1], so a client that connects by another name, such as the host's LAN address or a compose service name, needs that name in MCP_HTTP_ALLOWED_HOSTS. See HTTP transport security.
Option C: Install from source
git clone https://github.com/amittell/firewalla-mcp-server.git
cd firewalla-mcp-server
npm install
npm run build
2. Configuration
Create a .env file with your Firewalla credentials:
# Required
FIREWALLA_MSP_TOKEN=your_msp_access_token_here
FIREWALLA_MSP_ID=yourdomain.firewalla.net
# Optional - filters all queries to a specific box
# FIREWALLA_BOX_ID=your_box_gid_here
# Optional - default box for single-box operations, without filtering queries
# FIREWALLA_DEFAULT_BOX_ID=your_box_gid_here
# Optional - register the 11 write tools (default: off, read-only)
# FIREWALLA_ENABLE_WRITE_TOOLS=true
Getting Your Credentials:
Log into your Firewalla MSP portal at https://yourdomain.firewalla.net
Your MSP ID is the full domain (e.g., company123.firewalla.net)
Generate an access token in API settings
(Optional) Find your Box GID in device settings to filter queries to a specific box, or retrieve available boxes using the get_boxes tool
Box ID is optional. Without FIREWALLA_BOX_ID, queries cover every box on the account. The few operations that act on one box (get_specific_alarm, archive_alarm, mute_alarm, delete_alarm, create_rule, rename_device) take a gid argument, and without one they use FIREWALLA_BOX_ID, then FIREWALLA_DEFAULT_BOX_ID, then the account's only box. On an account with several boxes and none of those set, get_specific_alarm, archive_alarm, mute_alarm and delete_alarm check each box, and create_rule and rename_device refuse and list the boxes.
Test mode.MCP_TEST_MODE=true starts the server without credentials, to check that it starts: it uses a dummy token, API (https://test.firewalla.net) and box instead of your settings, always runs on stdio, and cannot read your Firewalla data. With NODE_ENV=production the server refuses test mode and exits with code 1. The Docker image sets NODE_ENV=production, so pass -e NODE_ENV=development along with -e MCP_TEST_MODE=true.
Transport Configuration
The MCP server supports two transport modes:
Stdio Transport (Default): Standard input/output communication for Claude Desktop and similar MCP clients
MCP_TRANSPORT=stdio
HTTP Transport: HTTP server mode for Docker containers, MCP orchestrators, and external access
MCP_TRANSPORT=http
MCP_HTTP_PORT=3000 # Default: 3000
MCP_HTTP_PATH=/mcp # Default: /mcp. Other paths get 404
MCP_HTTP_HOST=127.0.0.1 # Address to listen on, no port (::1 or [::1] for IPv6). Default: 127.0.0.1 (0.0.0.0 in the Docker image)
MCP_HTTP_BEARER_TOKEN= # Clients must send Authorization: Bearer <token>. Required, 16+ characters, unless MCP_HTTP_HOST is loopback
MCP_HTTP_ALLOW_NO_TOKEN= # "true" starts beyond loopback without a token, for a network no untrusted machine can reach
MCP_HTTP_ALLOWED_HOSTS= # More Host header names to accept, comma-separated
MCP_HTTP_ALLOWED_ORIGINS= # Browser origins to accept, comma-separated, e.g. http://localhost:6274
HTTP transport security
Every request can spend your MSP token, so the HTTP server follows the security rules of the MCP transport specification:
It listens on 127.0.0.1, this machine only. Set MCP_HTTP_HOST=0.0.0.0 (or one address) to accept other machines, and set MCP_HTTP_BEARER_TOKEN with it: on any address but loopback (127.0.0.0/8, ::1, localhost) the server refuses to start without a token, and exits with code 1 after one line on stderr that says why. MCP_HTTP_ALLOW_NO_TOKEN=true starts it without one, for a network no untrusted machine can reach, such as a compose network with no published port; the server then writes a warning to stderr at startup, whatever LOG_LEVEL says.
With MCP_HTTP_BEARER_TOKEN set, a request without Authorization: Bearer <token> gets 401. A token shorter than 16 characters stops startup wherever the server listens; openssl rand -hex 32 makes one 64 characters long.
A request whose Host header is not localhost, 127.0.0.1, [::1], the MCP_HTTP_HOST address (unless that is the wildcard 0.0.0.0 or ::, which adds nothing) or a name in MCP_HTTP_ALLOWED_HOSTS gets 403. This stops DNS rebinding, where a web page points its own domain name at your machine.
A request with an Origin header, which browsers send, gets 403 unless the origin is in MCP_HTTP_ALLOWED_ORIGINS, compared as URL origins (letter case and a default port make no difference). Non-browser MCP clients send no Origin and are not affected. An allowed origin gets CORS headers on every answer, refusals such as a 401 for a missing token included, so a web page on it can call the server and read why a request was refused.
The MCP endpoint is the configured MCP_HTTP_PATH exactly (default /mcp), that path with one trailing slash, and either with a query string. Any other path gets 404, CORS preflight requests included: with the default, /mcpx and /mcp/x do.
A request body may be at most 1 MB, and a client has 10 seconds to send the headers and 30 seconds for the whole request. An answer given without reading the body (401, 403, 404, 405, a malformed session ID, a CORS preflight) closes the connection, so a client cannot hold one open by sending the body slowly.
A session ends when the client sends DELETE, after MCP_SESSION_IDLE_TIMEOUT_MS without a request (default 30 minutes), or when the server restarts. A request with the ID of a session the server does not hold gets 404, and the transport specification has the client start a new session with a new initialize. A request with no session ID, other than initialize, gets 400.
When to use HTTP transport:
Running in Docker containers independently
Accessing from MCP orchestrators (e.g., open-webui)
Multiple clients need to connect to the same server instance
Network-based access to the MCP server
When to use stdio transport:
Claude Desktop integration (default)
Claude Code CLI integration
Single-process MCP client setups
Standard MCP client configurations
3. Build and Start
npm run build
npm run mcp:start
4. Connect Claude Desktop
Add this configuration to your Claude Desktop claude_desktop_config.json:
Review client-specific setup guides in docs/clients/
Usage Examples
Step-by-Step First Use
1. Verify Connection
After completing the setup, verify the MCP server is working:
# Start the server
npm run mcp:start
# stdout carries only the MCP protocol; stderr shows JSON log lines, the
# last one with "message":"Firewalla MCP Server running on stdio transport"
2. Test with Claude
Open Claude Desktop and try these starter queries:
Basic Health Check:
"Can you check my Firewalla status and show me a summary?"
This uses: the firewalla://summary resource and the get_simple_statistics tool
Security Overview:
"What security alerts do I have? Show me the 5 most recent ones."
This uses: get_active_alarms tool with limit parameter
Practical Workflows
Daily Security Review:
"Give me today's security report. Include:
1. Any new security alerts
2. Top 3 devices using bandwidth
3. Any devices that went offline
4. Status of critical firewall rules"
Investigating Suspicious Activity:
"I noticed unusual traffic. Can you:
1. Show me all security and abnormal upload alarms from the last 4 hours
2. Find any blocked connections to external IPs
3. Check which devices had the most network activity"
Network Troubleshooting:
"A device seems to have connectivity issues. Can you:
1. Check if device 192.168.1.100 is online
2. Show its recent network flows
3. See if any rules are blocking its traffic"
Bandwidth Investigation:
"Our internet is slow. Help me find the cause:
1. Show top 10 bandwidth users in the last hour
2. Look for any devices with unusual upload/download patterns
3. Check for any streaming or video traffic"
Advanced Search Examples
Find Specific Threats:
search for: security activity alarms from IP range 10.0.0.* in the last 24 hours
Uses: search_alarms with query: "type:1 AND device.ip:10.0.0. AND ts:>=<unix time 24 hours ago>"*
Analyze Rule Effectiveness:
"Show me firewall rules that blocked the most connections this week"
Uses: get_network_rules + search_flows for blocked traffic analysis
Device Behavior Analysis:
"Find all devices that were online yesterday but are offline now"
Uses: get_offline_devices, which lists offline devices most recently seen first, with last_seen; search_devices takes no time fields
Troubleshooting Common Issues
Connection Problems:
If you get authentication errors:
Verify your .env file has correct credentials
Check your MSP token hasn't expired
Confirm your Box ID is the full GID format
Empty Results:
If queries return no data:
Check your Firewalla is online and reporting
Verify the time range isn't too narrow
Try broader search terms first
Performance Issues:
If responses are slow:
Reduce the limit parameter in queries
Use more specific time ranges
Check your network connection to the MSP API
Available Tools (28 read-only, 11 opt-in write tools)
Every read-only tool (readOnlyHint: true) takes an optional response_format. json, the default, returns the compact JSON response. markdown returns the same response as markdown for reading: a heading with the tool name and the number of records in each list, the fields as bullet lists, and each list of records as a table of at most 8 columns and DEFAULT_PAGE_SIZE rows (default 100). Fields with one value in every record are stated once above the table, fields that are not columns are named below it, and a list of one record is shown as a list of its fields. The last line says what the view leaves out (the response's meta block, rows past the cap, cells cut at 120 characters); response_format: json returns all of it. Values and field names are escaped, since device names, domains and alarm messages come from the network: a character that could start markdown or HTML ([, <, *, `, & before an entity, ://, @, ...) gets a backslash, so a renderer shows [a](https://...) or <img ...> as text. Addresses, MACs, gids and timestamps read as they are. Omitting response_format or sending null means json. An error is compact JSON in either format: only a successful read is rendered as markdown. The tools that change state do not take response_format.
get_device_status with {"limit": 2, "response_format": "markdown"}, shortened:
## get_device_status (devices: 2)
- **total_devices:** 2
- **online_devices:** 1
- **devices:** 2 records, under devices below
### devices (2 records)
Same in every record: gid = 00000000-0000-0000-0000-000000000000; ip_reserved = false.
| name | id | last_seen | ip | online |
| --- | --- | --- | --- | --- |
| nas | AA:BB:CC:DD:EE:01 | 2026-09-26T10:00:00.000Z | 192.168.1.10 | true |
| printer | AA:BB:CC:DD:EE:02 | 2026-09-25T18:30:00.000Z | 192.168.1.11 | false |
_This view leaves out the meta block (request_id req_1790000000000_abc123). Call again with `response_format: json` for the full JSON response._
Write tools (opt-in)
Every tool that changes something is off by default, so the server is read-only unless you turn them on. Set FIREWALLA_ENABLE_WRITE_TOOLS=true to register the 11 write tools: create_rule, delete_rule, pause_rule and resume_rule (rules), create_target_list, update_target_list and delete_target_list (target lists), rename_device (devices), and archive_alarm, mute_alarm and delete_alarm (alarms). Without it, calling one answers "Unknown tool" and sends nothing. MCP clients that honor tool annotations can ask before calling them; see Tool annotations.
Up to 1.5.0, pause_rule, resume_rule and the three target-list tools were always registered. If you use them, set FIREWALLA_ENABLE_WRITE_TOOLS=true; see Upgrading to 2.0.0.
The write tools need an MSP API token with write access. Firewalla said on 2026-09-08 that MSP 2.12 adds read-only API tokens, which cannot make changes. When a write gets HTTP 403, the error says the token may be read-only and that the write tools need a token with write access. A 403 can also mean the request named a box the token cannot access, and the error says that too; get_boxes lists the boxes the token can access.
IDs that go into a request path (id, rule_id, alarm_id, gid, device_id) are checked before anything is sent. One holding /, a backslash, ?, #, %, whitespace or a control character, or that is . or .., is refused as a validation error naming the argument; leading or trailing whitespace is refused too, not trimmed. Rule IDs such as <box gid>:<n>, MAC device IDs and ovpn: device IDs are accepted.
create_rule and rename_device act on one box: gid, else FIREWALLA_BOX_ID or FIREWALLA_DEFAULT_BOX_ID, else the account's only box. On a multi-box account with none of those, they refuse without writing anything, because the MSP API applies a rule with no gid to every box in the account, including boxes added later. delete_rule needs MSP 2.11.0 or later.
archive_alarm and mute_alarm need MSP 2.11.0 or later. archive_alarm takes an alarm out of the active alarms and does nothing else: future matching traffic can still raise new alarms. mute_alarm archives the alarm and has the box create a lasting silence exception. Its target_type says what is silenced: alarmType every future alarm of that alarm's type, whatever the destination; domain a domain and its subdomains; ip one address. Its scope_type says for which devices: all of them, or the one device, group, user or network named by scope_value. The mute is checked against the documented request model before anything is sent, and this server has no tool to remove the exception afterwards. delete_alarm deletes the alarm for good (measured on 2026-09-26: the alarm was gone afterwards); use archive_alarm to keep it among the archived alarms. Alarm IDs are per box, and the same ID can name different alarms on different boxes: all three tools use the alarm's gid, else FIREWALLA_BOX_ID, else check each box, and they refuse when several boxes have that alarm ID and none of them is FIREWALLA_DEFAULT_BOX_ID. All three read the alarm before writing, so a wrong ID fails without changing anything.
Tool annotations
Every tool carries MCP tool annotations: a title, readOnlyHint, and openWorldHint: true, since each one calls the Firewalla MSP API. All get_* and search_* tools are read-only. The tools that change state have readOnlyHint: false, and all of them need FIREWALLA_ENABLE_WRITE_TOOLS=true:
Tool
destructiveHint
idempotentHint
pause_rule, resume_rule
false
true
create_target_list
false
false
update_target_list, delete_target_list
true
true
create_rule
true
false
delete_rule
true
true
rename_device
false
true
archive_alarm
false
true
mute_alarm
true
false
delete_alarm
true
true
pause_rule and resume_rule check the rule's status first and change nothing when it is already paused or active. Clients that honor annotations can ask before calling any tool that is not read-only.
Development
Scripts
npm run dev # Start development server with hot reload
npm run build # Build TypeScript to JavaScript
npm run test # Run all tests
npm run test:watch # Run tests in watch mode
npm run lint # Run ESLint
npm run lint:fix # Fix ESLint issues
MCP Execution Methods
Why npx for MCP servers?
Version Management: Always uses the correct/latest version
No global installation required: Works without global installation
MCP Standard: Follows Model Context Protocol conventions
Reliable: Works consistently across different environments
Alternative execution methods:
# Development (from source)
npm run mcp:start
# Production (npm installed)
npx firewalla-mcp-server
# Direct execution (from source after build)
node dist/server.js
Launch checks in CI
scripts/launch-smoke.mjs starts the server the way a user would and requires an answer to an MCP initialize over stdio. It uses dummy credentials, and initialize is answered locally, so nothing is sent to Firewalla.
The CI workflow's launch job packs the package and starts it through the global bin, npx and node dist/server.js: on Linux for a pull request, and on Linux, macOS and Windows for a push to main or a run by hand (Actions, CI, Run workflow). A change to docs/, top-level Markdown files or LICENSE alone runs no CI; no job reads them (format:check covers src/ only).
The Docker Build workflow runs on pull requests that touch the Dockerfile, the package files or the Docker workflows. It builds the image for amd64, arm64 and arm/v7, then runs the amd64 image with docker run -i --rm and requires serverInfo.version to equal package.json's version. Run the workflow by hand with the image input (for example amittell/firewalla-mcp-server:2.1.1) to pull and check a published image instead.
The Docker Build and Publish workflow, on a release tag, pushes the image and sets its Docker Hub description from DOCKERHUB.md, a short page that links here for the rest (Docker Hub keeps only the first 25,000 characters of a description). Run it by hand with description_only to update just the description.
To run the Docker check locally:
docker build -t firewalla-mcp-server:local .
node scripts/launch-smoke.mjs --docker firewalla-mcp-server:local
# A published image; the expected version comes from the tag
node scripts/launch-smoke.mjs --docker amittell/firewalla-mcp-server:2.1.1 --pull
Project Structure
firewalla-mcp-server/
├── src/
│ ├── server.ts # Main MCP server
│ ├── firewalla/ # Firewalla API client
│ ├── tools/ # MCP tool implementations
│ ├── resources/ # MCP resource implementations
│ └── prompts/ # MCP prompt implementations
├── tests/ # Test files
├── docs/
│ └── firewalla-api-reference.md # API documentation
├── CLAUDE.md # Comprehensive development guide
├── SPEC.md # Technical specifications
└── README.md # This file
MSP tokens are stored securely in environment variables
No credentials are logged or stored in code
Each server process paces its own requests to API_RATE_LIMIT per rolling 5 minutes (default 100, the limit measured on one MSP token; whether it is per token or per account was not measured). A request that cannot start within 20 s fails at once, saying when capacity returns. Every MCP client that starts the server over stdio starts its own process with its own count, so several of them on one token need API_RATE_LIMIT values that add up to 100 or less. See the rate-limiting guide
Input validation prevents injection attacks
Device names, domains and alarm messages are set by the devices and sites on your network, so the server treats them as untrusted: characters that do not display are shown as markers such as <U+E0041>, the prompts quote API data only inside a marked data block, and the initialize instructions tell the client to treat results as data. See Untrusted data
All API communications use HTTPS
Known Behaviors and Limitations
Category Classification
Flow Categories: Many network flows may show as empty category ("") in the Firewalla API response. This is expected behavior - Firewalla categorizes traffic when it recognizes the domain/service (e.g., "av" for audio/video, "social" for social media).
Target List Categories: Some target lists may show category as "unknown". This is normal for user-created or certain system lists.
Timeline: Category classification happens at the Firewalla device level and may take time to build up meaningful categorization data.
Data Characteristics
Response Sizes: The get_recent_flow_activity tool reads the 50 most recent flows in one request, so it covers minutes on a busy network. For larger datasets or historical analysis, use search_flows with time filters for more targeted queries.
Geographic Data: IP geolocation is enriched by the MCP server and includes country, city, and risk scores when available.
API Limitations
MSP only: The server sends every request to the MSP API at https://<FIREWALLA_MSP_ID> and never connects to the box on your LAN. It has no local mode, so it needs an MSP account with API access even for one box at home, and it sees only what the MSP API returns.
Alarm Deletion: In July 2025 the MSP API answered DELETE /v2/alarms/{gid}/{aid} with {"message": "success", "success": true} and kept the alarm, so delete_alarm was withdrawn. Measured again on 2026-09-26, the same request deleted the alarm (a GET of it then answered 404, and the archived alarm count fell by one), and delete_alarm is back as an opt-in write tool.
Troubleshooting
Quick Fixes
Server won't start:
# Clean and rebuild
npm run clean
npm run build
# If build fails, try:
npm install
npm run build
Model Context Protocol (MCP) server for Firewalla MSP API - Provides real-time network monitoring, security analysis, and firewall management through 28 read-only tools, plus 11 opt-in write tools, co
Model Context Protocol (MCP) server for Firewalla MSP API - Provides real-time network monitoring, security analysis, and firewall management through 28 read-only tools, plus 11 opt-in write tools, compatible with any MCP client
上次檢查時間:2026-10-28。 Model Context Protocol (MCP) server for Firewalla MSP API - Provides real-time network monitoring, security analysis, and firewall management through 28 read-only tools, plus 11 opt-in write tools, compatible with any MCP client