NetGet

LLM-Controlled Network Protocol Server & Client

NetGet is a Rust CLI application exposing 283 network protocols, with server and client roles controlled by an LLM (via Ollama). Instead of hardcoding protocol logic, NetGet provides the network stack while the LLM constructs raw protocol datagrams or high-level responses based on natural language instructions.

# Start a MySQL server that reads schema from files
netget "act as MySQL server, read schema.json for database structure"

# Start an SSH server with custom authentication
netget "SSH server on port 2222, allow users alice and bob"

# Connect as HTTP client to test an API
netget "connect to https://api.example.com as HTTP client"

Why NetGet?

Traditional Approach: Hardcode protocol behavior

match command {
    "USER" => send("331 Password required\r\n"),
    "PASS" => send("230 Logged in\r\n"),
    // ... hundreds of lines of rigid logic
}

NetGet's Approach: Instruct the LLM in natural language

> act as an HTTP server, serve /data.txt with content 'hello', 404 everything else

The LLM handles the protocol details — status lines, headers, authentication exchanges — without any code changes.

Key Features

🌐 283 Network Protocols

Both server and client modes for:

Core Transport

• TCP, UDP, HTTP, HTTP/2, HTTP/3, TLS, DataLink

Application Protocols

• SSH, FTP, DNS, DHCP, SMTP, IMAP, MySQL, PostgreSQL, Redis
• MQTT, Kafka, gRPC, WebSocket, WebDAV, NFS, SMB
• WireGuard, Tor; IPSec and OpenVPN are honeypots, not working VPNs

Maturity varies a lot. Most protocols are Experimental — LLM-authored and not human-reviewed. Run netget --mcp and call list_protocols for each protocol's current state, or read its metadata(). A few specifics worth knowing: FTP has no data connection, so LIST/RETR/STOR cannot complete against a real client; OpenVPN is a control-plane honeypot — a real openvpn binary completes the TLS control channel and the key method 2 exchange, but PUSH_REQUEST is never answered, no data-channel keys are derived and there is no TUN device, so it observes clients and never carries traffic.

IoT & Hardware

• Bluetooth Low Energy (15+ GATT services: keyboard, mouse, heart rate, etc.)
• USB/IP (keyboard, mouse, serial, mass storage, FIDO2, smart card)

Routing & Infrastructure

• BGP, OSPF, IS-IS, RIP, mDNS, SNMP, NTP, Syslog

Specialized

• OAuth2, SAML, OpenID Connect, MCP, JSON-RPC, XML-RPC
• BitTorrent (tracker, DHT, peer), OpenAPI, Kubernetes, etcd

See the /docs command in the TUI for full protocol details and metadata.

🤖 LLM-Driven Intelligence

• Tool Calling: LLM can read files and search the web before responding
• Scripting Mode: Use Python/JavaScript for deterministic protocol logic (0 LLM calls)
• Scheduled Tasks: Time-based automation at global, server, or connection scope
• Memory: LLM maintains conversation history across requests

🖥️ Rich Terminal Interface

Rolling Terminal UI with sticky footer:

• Three-column layout: Active servers | Active clients | Status & stats
• Natural scrolling: Output flows into your terminal's scrollback buffer
• Multi-line input: Shift+Enter for complex commands
• Command history: Up/down arrows with persistent storage
• Log levels: Ctrl+L to toggle ERROR/WARN/INFO/DEBUG/TRACE
• Web search: Ctrl+W to enable LLM web search capability

Quick Start

Prerequisites

1. Install Ollama (LLM runtime)

   curl https://ollama.ai/install.sh | sh
   ollama pull qwen3-coder:30b  # or another model
   

2. Install Rust (latest stable)

   curl --proto '=https' --tlsv1.2 -sSf https://sh.rustup.rs | sh
   

Installation

Option 1: npm (prebuilt binary, no Rust needed)

# Run directly (downloads the binary for your platform on first use)
npx @smotana/netget

# Or install globally
npm i -g @smotana/netget
netget

Prebuilt binaries cover macOS (arm64/x64), Linux (x64/arm64 glibc, x64 musl), and Windows x64. They need no system library: on Linux and Windows they leave out the few protocols that link one (packet capture, Bluetooth LE, the NFC client, USB devices, the SMB client); macOS ships those libraries, so its binaries include them. Tarballs are also on GitHub Releases.

Option 2: cargo (every protocol)

The crate is netget-tool (crates.io's netget is an unrelated project); the command it installs is netget. Every protocol is compiled in — there are no per-protocol features. The default build links five system libraries and builds llama.cpp (which needs cmake and libclang), so install those first:

# macOS (libpcap, PC/SC and CoreBluetooth ship with the OS)
brew install samba libusb pkgconf cmake

# Debian / Ubuntu
sudo apt-get install build-essential pkg-config cmake libclang-dev libssl-dev libltdl-dev \
    libpcap-dev libdbus-1-dev libusb-1.0-0-dev libpcsclite-dev libsmbclient-dev

# Fedora
sudo dnf install gcc gcc-c++ make pkgconf-pkg-config cmake clang-devel openssl-devel libtool-ltdl-devel \
    libpcap-devel dbus-devel libusb1-devel pcsc-lite-devel libsmbclient-devel

# Arch
sudo pacman -S --needed base-devel cmake clang openssl libpcap dbus libusb pcsclite smbclient

cargo install --locked netget-tool

Without those libraries, leave out the protocols that need them (and the embedded llama.cpp):

cargo install --locked netget-tool --no-default-features --features mcp-stdio,mcp-http,sqlite

Option 3: Build from source

git clone https://github.com/netget-net/netget.git
cd netget
cargo build --release          # same system libraries as Option 2

Run Your First Server

# Start Ollama
ollama serve

# Start NetGet
./target/release/netget

# In the NetGet prompt, type:
> listen on port 8080 via HTTP, serve a page that says "Hello World"

Now visit http://localhost:8080 in your browser!

Architecture

NetGet uses an event-driven architecture where all network events flow through a central event handler that coordinates with the LLM:

┌─────────────────────────────────────────────────────────┐
│                   Rolling Terminal UI                    │
│  ┌──────────┐  ┌──────────┐  ┌──────────────────────┐  │
│  │  Active  │  │  Active  │  │  Connection Info &   │  │
│  │  Servers │  │  Clients │  │  Stats               │  │
│  └──────────┘  └──────────┘  └──────────────────────┘  │
│  ┌──────────────────────────────────────────────────┐  │
│  │         Status / Activity Log (Scrolling)        │  │
│  └──────────────────────────────────────────────────┘  │
│  ┌──────────────────────────────────────────────────┐  │
│  │  Input: > your commands here (Shift+Enter)       │  │
│  └──────────────────────────────────────────────────┘  │
└─────────────────────────────────────────────────────────┘
                            │
         ┌──────────────────┴──────────────────┐
         ▼                                     ▼
┌──────────────────┐               ┌──────────────────────┐
│  Network Layer   │               │   LLM Integration    │
│  • TCP/UDP       │◄─────────────►│  • Ollama Client     │
│  • TLS           │               │  • Tool Calling      │
│  • Raw Sockets   │               │  • Scripting         │
│  • USB/IP        │               │  • Web Search        │
│  • Bluetooth     │               │                      │
└──────────────────┘               └──────────────────────┘

Key Modules

• cli/ - Terminal UI (rolling output, sticky footer, input handling)
• server/<protocol>/ - Protocol server implementations
• client/<protocol>/ - Protocol client implementations
• protocol/ - Protocol registry and metadata
• llm/ - Ollama integration and prompt engineering
• llm/actions/ - LLM action system (user-triggered and network-triggered)
• events/ - Event coordination between network, LLM, and UI
• state/ - Application state management

Design Principles

1. Decentralized Protocol Registry: Each protocol implements traits independently (no central switch statements except in startup)
2. Dual Logging: All logs go to both tracing macros (→ netget.log) and status channel (→ TUI)
3. State Machine: Per-connection state (Idle → Processing → Accumulating) prevents concurrent LLM calls
4. Every Protocol In Every Build: No per-protocol Cargo features; only the platform (browser, Windows) and five system-library features decide what is compiled
5. No Protocol Storage: Protocols don't store data—LLM returns everything via actions/scripts/static responses

Usage Examples

Server Examples

HTTP Server with Dynamic Content

> listen on port 8080 via HTTP
> For /api/time return JSON with current timestamp
> For /api/random return a random number between 1 and 100

MySQL Server with Schema File

# Create schema file first
$ cat > schema.json <<EOF
{
  "database": "shop",
  "tables": [
    {"name": "products", "columns": ["id", "name", "price"]},
    {"name": "users", "columns": ["id", "email", "created_at"]}
  ]
}
EOF

> act as MySQL server, read schema.json for database structure
> Answer SELECT queries based on the schema

SSH Server with Custom Behavior

> SSH server on port 2222
> Allow password authentication for user "admin" with password "secret"
> When user runs "date", return current date and time
> When user runs "fortune", return a random fortune cookie message

DNS Server with Dynamic Responses

> DNS server on port 5353
> For *.local domains, return 127.0.0.1
> For *.dev domains, return 192.168.1.100
> For google.com, return 8.8.8.8

Client Examples

HTTP Client Testing API

> connect to https://api.github.com as HTTP client
> Send GET request to /users/octocat
> Show me the response headers and body

Redis Client Session

> connect to 127.0.0.1:6379 as Redis client
> Send: SET mykey "Hello World"
> Send: GET mykey
> Send: INCR counter

TCP Client for Protocol Testing

> connect to localhost:25 as TCP client
> This is an SMTP server, try sending an email

Advanced Features

Tool Calling (File Reading)

> act as HTTP server on port 8000
> For /config, read config.json and return it as JSON
> For /users, read users.csv and return it formatted as HTML table

The LLM will automatically call read_file() when handling requests.

Web Search Integration

# Enable web search
Ctrl+W

> act as HTTP server on port 80
> If you receive a request you don't understand, search for the HTTP RFC

The LLM can search the web for protocol documentation, RFCs, or any other information.

Scripting Mode (Zero LLM Calls)

> listen on port 3000 via HTTP
> Use JavaScript scripting mode for responses
> For /api/time return: { "time": new Date().toISOString() }
> For /health return: { "status": "ok" }

Scripting mode is deterministic and makes zero LLM calls after setup.

Scheduled Tasks

> listen on port 22 via SSH
> Schedule a task every 60 seconds to broadcast "System health check complete"
> Schedule a task after 300 seconds to close all idle connections

Tasks can be scoped to global (any server), server-specific, or connection-specific.

Managing Connections

> status                    # Show all active servers and clients
> close server 1            # Close specific server
> close client 2            # Close specific client
> close all                 # Close everything
> model llama3.3:70b        # Switch to different LLM model

Control NetGet from Another LLM (MCP)

NetGet can run as a Model Context Protocol server, exposing its protocol-server management as MCP tools. This lets an MCP client such as Claude Desktop or Claude Code start and drive network protocol servers directly.

STDIO transport (Claude Desktop / Claude Code)

Quickest path — npm prebuilt binary:

claude mcp add netget -- npx -y @smotana/netget --mcp

Or build from source (mcp-stdio is a default feature):

./cargo-isolated.sh build --release

# Run as an MCP server over stdin/stdout
netget --mcp

Claude Desktop config (claude_desktop_config.json):

{
  "mcpServers": {
    "netget": {
      "command": "npx",
      "args": ["-y", "@smotana/netget", "--mcp", "--model", "qwen3-coder:30b"]
    }
  }
}

(For a source build, set "command" to the binary path instead. First npx run downloads the package, so the initial MCP startup takes longer.)

HTTP transport (remote / web clients)

# mcp-http is a default feature
./cargo-isolated.sh build --release

# Serve MCP over HTTP/SSE (bind address from --listen-addr, default 127.0.0.1)
netget --mcp-http 8080          # endpoint: http://127.0.0.1:8080/mcp

# Any non-loopback bind requires a bearer token
NETGET_MCP_TOKEN=$(openssl rand -hex 32) netget --mcp-http 8080 --listen-addr 0.0.0.0

With a token set — NETGET_MCP_TOKEN, or --mcp-token <TOKEN> — every request must carry Authorization: Bearer <TOKEN>. Binding --listen-addr to anything but a loopback address is refused without one; without a token the endpoint admits loopback Host/Origin only. Prefer the environment variable: a --mcp-token argument is visible to every local user in the process table.

Exposed tools

list_protocols, start_server, stop_server, list_servers, server_status, get_status, set_model, get_protocol_docs, update_server_instruction, stop_all, and for debugging: list_access_logs (recent requests) / get_access_log (full request + response for one entry).

The protocol servers you start through MCP are driven by NetGet's own configured LLM (local Ollama, or an OpenAI-compatible endpoint via --openai-url/--api-key). The MCP client controls NetGet through the tools above; it does not supply the model.

Testing

NetGet has comprehensive test coverage with both unit tests and end-to-end tests.

Unit Tests (No Ollama Required)

./cargo-isolated.sh test --lib

Tests protocol parsing, registry, metadata, and core utilities.

E2E Tests (Ollama Required)

# Start Ollama first
ollama serve
ollama pull qwen3-coder:30b

# Run one protocol's E2E tests (every protocol is compiled in; select by module path)
./cargo-isolated.sh test --test server -- server::tcp::e2e_test

# Run all of a protocol's tests (server + client)
./cargo-isolated.sh test --test server --test client -- server::http:: client::http::

E2E Test Philosophy:

• Black-box testing: Tests use real clients (curl, mysql CLI, ssh clients, etc.)
• LLM budget: Each test suite limited to < 10 LLM calls via aggressive caching
• Prompt-driven: Tests validate that LLM interprets prompts correctly
• Localhost only: All tests run on 127.0.0.1/::1 for privacy

Test Organization

• Unit tests: tests/base_stack_test.rs (registry parsing, etc.)
• Protocol E2E: tests/server/<protocol>/e2e_test.rs, tests/client/<protocol>/e2e_test.rs
• Documentation: Each protocol has TWO CLAUDE.md files:
  • src/server/<protocol>/CLAUDE.md - Implementation details
  • tests/server/<protocol>/CLAUDE.md - Test strategy

Building

System Dependencies

Every protocol is compiled into every build; there are no per-protocol Cargo features. Five features link a system library (all on by default, Unix only), and embedded-llm builds llama.cpp:

The install commands for each platform are under Installation above.

Build Strategies

# Everything (needs the libraries above)
./cargo-isolated.sh build

# Everything that needs no system library
./cargo-isolated.sh build --no-default-features --features mcp-stdio,mcp-http,sqlite

Claude Code for Web has none of the five libraries:

./am_i_claude_code_for_web.sh
./cargo-isolated.sh build --no-default-features --features mcp-stdio,mcp-http,sqlite

Build Performance Notes

• sccache: NetGet uses sccache for caching compiled artifacts
• No per-protocol features: every build compiles every protocol; narrow test runs with filters (--test server -- server::tcp::), not features
• Isolated builds: ./cargo-isolated.sh uses session-specific target directory to avoid conflicts
• Parallel builds: Multiple instances can build concurrently; ./cargo-isolated.sh serialises access to the shared target/

Cargo Isolated Scripts

./cargo-isolated.sh build                     # Build with isolated target dir
./cargo-isolated.sh test --test server -- server::http::   # Test one protocol
./cargo-isolated.sh run -- --debug            # Run with debug logging
./cargo-isolated.sh --print-last              # View last build/test output
./cargo-isolated-kill.sh                      # Kill all isolated builds (not system cargo!)

Output is automatically logged to ./tmp/netget-<command>-{PPID}.log.

Configuration

LLM Model Selection

Change at runtime (recommended):

> model qwen3-coder:30b
> model llama3.3:70b
> model codestral:latest

Default model: qwen3-coder:30b (configured in src/state/app_state.rs)

Ollama URL

Default: http://localhost:11434

To change, edit src/llm/client.rs:

pub fn default() -> Self {
    Self::new("http://your-ollama-host:11434")
}

Log Levels

• Runtime: Press Ctrl+L to cycle through ERROR → WARN → INFO → DEBUG → TRACE
• Startup: netget --debug enables debug logging to netget.log

History

Command history is persisted to ~/.netget_history across sessions.

Protocol Implementation

Adding a New Server Protocol

See CLAUDE.md for the full 12-step checklist. Quick overview:

1. Add feature to Cargo.toml
2. Implement server in src/server/<protocol>/mod.rs
3. Implement actions in src/server/<protocol>/actions.rs
4. Register in protocol/registry.rs (feature-gated)
5. Add match arm in cli/server_startup.rs (feature-gated)
6. Write E2E test in tests/server/<protocol>/e2e_test.rs
7. Document in src/server/<protocol>/CLAUDE.md and tests/server/<protocol>/CLAUDE.md

Adding a New Client Protocol

Similar to server, but:

• Implement in src/client/<protocol>/mod.rs and actions.rs
• Register in protocol/client_registry.rs
• Add match arm in cli/client_startup.rs
• See CLIENT_PROTOCOL_FEASIBILITY.md for protocol evaluation criteria

Advanced Topics

Multi-Instance Collaboration

• Use ./cargo-isolated.sh for per-session build isolation
• Use --llm-max-concurrent (with --llm-queue-timeout and --llm-max-queued) to bound LLM concurrency
• Use git worktrees for concurrent development branches
• Never use pkill cargo—use ./cargo-isolated-kill.sh instead

Efficient Iteration

Build/test logging is automatic. Analyze all errors before rebuilding:

# Build and pipe last 50 lines
./cargo-isolated.sh build | tail -50

# Analyze ALL errors from saved log
./cargo-isolated.sh --print-last | grep "error\[E"

# Fix ALL issues at once, then rebuild once
./cargo-isolated.sh build | tail -50

This saves 10-20 minutes per development cycle vs fixing one error at a time.

Protocol-Specific Documentation

Each protocol has detailed documentation:

• Implementation: src/server/<protocol>/CLAUDE.md or src/client/<protocol>/CLAUDE.md
  • Library choices, architecture, LLM integration, limitations
• Testing: tests/server/<protocol>/CLAUDE.md or tests/client/<protocol>/CLAUDE.md
  • Test strategy, LLM call budget, runtime, known issues

Always read both before modifying a protocol.

Troubleshooting

Ollama Connection Failed

Error: Failed to connect to Ollama

Fix: Ensure Ollama is running: ollama serve

Model Not Found

Error: Model not found

Fix: Pull the model: ollama pull qwen3-coder:30b

Port Already in Use

Error: Address already in use

Fix: Choose a different port or kill the process using that port.

Build Fails with Missing Libraries

error: failed to run custom build command for `dbus`

Fix: Install the system libraries (see Installation), or build without them: --no-default-features --features mcp-stdio,mcp-http,sqlite leaves out only the protocols that link one.

Performance Considerations

• LLM latency: Each network event may trigger an LLM call (can be slow)
• Scripting mode: For deterministic protocols, use Python/JavaScript scripting (zero LLM calls)
• Caching: Connection state machine prevents concurrent LLM calls on same connection
• LLM concurrency: --llm-max-concurrent bounds in-flight model calls; requests above it queue rather than being dropped

Use case: NetGet is ideal for protocol testing, security research, learning, and development. Not recommended for production high-throughput servers.

Contributing

NetGet is an experimental project exploring LLM-controlled networking. Contributions welcome!

Ideas for enhancement:

1. New protocol implementations
2. Better prompt engineering for existing protocols
3. Scripting mode templates for common protocols
4. Performance optimizations (caching, connection pooling)
5. Additional tool calling capabilities
6. Protocol fuzzing and security testing features

Files Created

NetGet creates the following files:

• netget.log - Application logs (only with --debug flag)
• ~/.netget_history - Command history (always created)
• ./tmp/netget-*.log - Build/test logs from cargo-isolated.sh

License

MIT

Acknowledgments

• Built with Tokio for async runtime
• Terminal UI powered by Crossterm
• LLM integration via Ollama
• 283 protocol features, built on best-in-class Rust crates