Sift (Docker Edition)

An MCP server that exposes the SANS SIFT digital-forensics toolkit as tools an LLM can call. Drive it from Claude Desktop, Ollama Desktop, or any MCP client.

This edition runs the whole toolchain as a self-contained Docker container — no SIFT VM required, works anywhere Docker runs. If you'd rather run on an existing SANS SIFT install, see the VM edition (separate repo): sift-mcp-vm.

It wraps standard DFIR command-line programs (The Sleuth Kit, Volatility 3, Plaso, exiftool, YARA, etc.) behind a safe, allowlisted interface. This is a defensive / investigative tool — it does not generate exploits or malware.

Contents

sift-mcp-docker/
├── server.py            # the MCP server
├── requirements.txt     # Python dependencies
├── Dockerfile           # Ubuntu 24.04 + forensic toolchain + server
├── docker-compose.yml   # service, mounts, ports, healthcheck
├── healthcheck.py       # container liveness probe
├── .env.example         # host bind address, port, timeouts
├── .dockerignore
├── clients/             # example client configs (Claude Desktop, Ollama)
└── README.md

The image is Ubuntu 24.04 with The Sleuth Kit, foremost, exiftool, binwalk, yara, and Plaso (from the GIFT PPA), plus Volatility 3 and python-evtx (via pip). It runs as a non-root user and serves MCP over HTTP on port 8000.

Quick start

git clone <your-repo-url> sift-mcp-docker
cd sift-mcp-docker
cp .env.example .env            # optional: tweak ports/limits
mkdir -p cases/output           # evidence goes in ./cases
docker compose up -d --build    # build + start
docker compose logs -f          # watch startup

The endpoint is now http://localhost:8000/mcp. Put evidence in ./cases (e.g. ./cases/case01/disk.E01); recovered/carved files and timelines appear in ./cases/output.

How the mounts work

• ./cases → mounted read-only at /cases; the container can never alter your original evidence.
• ./cases/output → mounted writable at /cases/output for results.

If the container can't write to cases/output (a bind-mount permission mismatch), either chmod 777 cases/output on the host, or add user: "${UID}:${GID}" to the service in docker-compose.yml and run UID=$(id -u) GID=$(id -g) docker compose up -d.

Without compose

docker build -t sift-mcp:latest .
docker run -d --name sift-mcp -p 127.0.0.1:8000:8000 \
  -v "$PWD/cases:/cases:ro" -v "$PWD/cases/output:/cases/output" \
  sift-mcp:latest

Common commands

docker compose ps             # status + health
docker compose logs -f        # logs
docker compose restart        # restart after editing .env
docker compose down           # stop and remove
docker compose up -d --build  # rebuild after changing server.py

The tools

Disk / file analysis — sift_disk_partitions (mmls), sift_image_info (img_stat), sift_filesystem_info (fsstat), sift_list_files (fls), sift_extract_file (icat, returns SHA-256), sift_carve_files (foremost), sift_file_type (file), sift_hash_file (md5/sha1/sha256).

Memory forensics — sift_volatility (any Volatility 3 plugin).

Timeline & artifacts — sift_create_timeline (log2timeline), sift_export_timeline (psort), sift_parse_evtx.

Metadata & strings — sift_exiftool, sift_strings, sift_binwalk, sift_hexdump, sift_yara_scan.

Housekeeping — sift_list_evidence, sift_server_info (shows which binaries are installed). Call sift_server_info after startup to confirm the toolchain inside the container.

Connecting clients

Both configs are in clients/. Use http://localhost:8000/mcp as the URL.

Claude Desktop

Claude Desktop speaks MCP over stdio, so bridge to the HTTP endpoint with mcp-remote (needs Node.js on the host). Edit %APPDATA%\Claude\claude_desktop_config.json (Windows) or ~/Library/Application Support/Claude/claude_desktop_config.json (macOS):

{
  "mcpServers": {
    "sift": {
      "command": "npx",
      "args": ["-y", "mcp-remote", "http://localhost:8000/mcp", "--transport", "http-only"]
    }
  }
}

Ollama Desktop

Use mcphost to bridge an Ollama model to MCP:

go install github.com/mark3labs/mcphost@latest
mcphost -m ollama:qwen2.5 --config clients/ollama_mcphost_config.example.json

Use a tool-calling model (e.g. qwen2.5, llama3.1).

Quick check

curl -i http://localhost:8000/mcp

A 406 Not Acceptable is expected and good — it means the server is up (it only accepts proper MCP POSTs, not bare GETs).

Configuration (environment variables)

Set host-side values in .env; container-internal paths are fixed by compose.

Variable	Default	Purpose
HOST_BIND	127.0.0.1	Host address to publish on (0.0.0.0 for LAN access)
HOST_PORT	8000	Host port mapped to the container
SIFT_TIMEOUT	600	Default per-command timeout (s)
SIFT_MAX_OUTPUT_CHARS	60000	Output truncation limit

Security model

• Each tool runs one fixed, allowlisted binary via exec — never a shell.
• File paths are resolved (symlinks included) and confined to /cases (read) and /cases/output (write); anything outside is rejected.
• Plugin names, carve types, inode addresses, and parser names are shape-validated. Every command runs under a timeout with truncated output.
• The server has no authentication — by default it publishes only on 127.0.0.1. If you set HOST_BIND=0.0.0.0, keep it on a trusted/private network, never the public internet.