Wireshark MCP

Give your AI assistant a packet analyzer.

Drop a .pcap file, ask questions in plain English — get answers backed by real tshark data.

English • 中文 • Docs • Changelog • Roadmap • Contributing

What is this?

An MCP server that wraps tshark (and optional Wireshark suite tools) into a structured analysis interface. Works with Claude Desktop, Claude Code, Cursor, VS Code, and 18+ other MCP clients.

You:    "Find all DNS queries going to suspicious domains in this capture."
Claude: [calls wireshark_extract_dns_queries → wireshark_detect_dns_tunnel]
        "Found repeated high-entropy DNS queries consistent with tunneling: ..."

Install

Prerequisites: Python 3.10+ and Wireshark with tshark on PATH.

Wireshark MCP 3.0 uses the stable MCP Python SDK 2.x line (mcp>=2.1.1,<3).

pip install wireshark-mcp
wireshark-mcp install   # choose from detected MCP clients

Restart your AI client — done.

Run wireshark-mcp doctor if anything looks off. See docs/manual-configuration.md for manual setup or platform-specific notes.

Quick Start

Point your AI client at a .pcap file and try:

Analyze capture.pcap using the Wireshark MCP tools.
Start with wireshark_open_file, then run wireshark_quick_analysis.
Use wireshark_aggregate for any capture-wide count or distribution.
Write findings to report.md.

Tools

52 tools, each backed by real tshark output — organized into categories:

One tool covers 20 protocols rather than 20 tools covering one each: wireshark_analyze_protocol takes a protocol argument (tls_handshakes, mqtt, modbus, s7comm, zigbee, wifi, rtp, kerberos, …) and applies the right fields and display filter for it. The field names are the point — s7comm.param.item.dbnum is not something a caller should have to guess, and a wrong guess returns an empty result that reads like a clean capture.

The server starts with only tshark required. Optional tools (capinfos, mergecap, editcap, dumpcap, text2pcap) are auto-detected and enable extra features when present.

Context cost

The tool list travels in the prompt prefix of every request your client sends, so its size is a fixed per-request cost. The default surface is ~22 KB — about 9 KB of parameter schema, 5 KB of descriptions, and 3 KB of read/write annotations — and it is byte-identical across restarts so clients can cache the prefix rather than re-reading it each session.

If your client never captures live traffic or writes pcaps, --profile advertises less:

wireshark-mcp serve --profile core

Runtime prompts and protocol recommendations respect the selected profile. Static guides may describe full-only workflows, but the server never recommends an excluded tool during capture discovery.

Tool results are bounded too, since a result stays in the conversation for the rest of the session. Output over 8000 characters is truncated head-and-tail with a marker, and the tool's offset / limit / display_filter parameters are the way to page through the rest. Raise or lower the ceiling with:

export WIRESHARK_MCP_MAX_RESULT_CHARS=16000

Every tool also declares whether it reads or writes, so clients can auto-approve the 41 read-only analysis tools and still prompt for the 11 that create files (live capture, merge, filter-save, editcap, text2pcap, frame extract, object export).

In 3.0, those 11 tools fail closed until WIRESHARK_MCP_ALLOWED_DIRS names existing directories. Remote HTTP/SSE binding also stays loopback-only unless --allow-insecure-http is explicitly supplied behind a trusted authenticated TLS proxy. See the 3.0 security migration guide.

Documentation

Development

pip install -e ".[dev]"
pytest tests/ -v
ruff check src/ tests/

See CONTRIBUTING.md for the full guide.