🐋 CyberRescue

Give Claude eyes and hands inside your broken Docker containers.

📺 Live Triage Demonstration (33s)

https://github.com/user-attachments/assets/0ee0f583-b8c1-4abe-9dd4-4b59ec25fd49

Related MCP server: MCP Development Server

Overview

A locally-hosted MCP (Model Context Protocol) server that gives Claude real tools to debug Docker containers — fetch logs, inspect memory/CPU, and run diagnostic commands inside a container, all from a chat with Claude Desktop. Instead of writing bespoke glue code for every diagnostic endpoint you want an AI agent to reach (logs API, stats API, exec API, each with its own auth, sanitization, and error handling), CyberRescue exposes them once through MCP: register the server, and any MCP-capable client gets all three capabilities with validation, retry, and output caps already handled.

Results

• Per-endpoint integration time cut from ~2 hours to under 30 minutes. Wiring an agent to a new container-diagnostic capability previously meant hand-rolling the client call, input validation, output truncation, and failure handling. With the MCP tool pattern established here (validate → semaphore-gated daemon call → retry with back-off → sanitized errors → capped output), adding an endpoint is a single decorated function.

• 42 passing unit tests (30 for the core MCP server/security/retry logic, 12 for the public demo backend) covering input validation, command-safety blocklists, retry behavior, and the public-demo allowlist — run on every push via GitHub Actions CI.

• Verified A-Grade Quality on Glama and listed on global MCP indexes.

Live Demo

(Link goes here once deployed — backend setup: infra/README.md; frontend: deploy web/ to Vercel with its root directory set to web/.) A public demo of the real tool running against 3 intentionally broken sandboxed containers (broken-flask, leaking-node, crashed-nginx). It's the same stream_container_logs / inspect_memory_dump / execute_isolated_script logic used by the local MCP server, exposed over a small FastAPI backend that's locked to those 3 containers with a fixed diagnostic-command menu (no arbitrary containers, no freeform shell input) — see SECURITY.md for the full threat model. The local MCP path (this README, below) and the public demo are two different deployment contexts sharing one core (src/cyberrescue/core.py).

Claude Desktop --stdio--> server.py ---\
                                        +--> core.py (docker calls) --> Docker daemon
   Vercel (Next.js) --HTTPS--> backend/app --/

What it does

CyberRescue exposes three tools to Claude:

• stream_container_logs — fetch stdout/stderr logs from a container by ID or name (tail, since-timestamp, keyword filter; 50KB hard cap with truncation flag).

• inspect_memory_dump — live CPU/memory snapshot via docker stats, plus top processes via ps aux --sort=-%mem.

• execute_isolated_script — run a shell command inside a container via docker exec, with input validation, a command blocklist, and a hard asyncio timeout.

Everything runs locally over stdio — no network ports, no cloud service, no API keys beyond what you already use for Claude Desktop.

Architecture

flowchart LR
    A[Claude Desktop / MCP client] <-- "JSON-RPC over stdio" --> B[CyberRescue<br/>FastMCP server]
    B --> C{Security layer}
    C --> D["validate_container_id()"]
    C --> E["check_command_safety()<br/>command blocklist"]
    C --> F[Semaphore<br/>max 4 concurrent daemon calls]
    F --> G["Retry with exponential back-off<br/>(read-only calls: logs, stats)"]
    G --> H[(Docker daemon<br/>via python-on-whales)]
    H --> I[docker logs / docker stats / docker exec]
    B -- "sanitized errors, 50KB output cap" --> A

Read-only telemetry calls (logs, stats) retry up to 3 times with exponential back-off (0.5s → 1s) on transient daemon failures; docker exec is deliberately never retried, since it isn't idempotent. Raw Docker exceptions are logged server-side and never leaked to the client — they can contain host socket paths and usernames.

Requirements

• macOS (Apple Silicon) or Windows 10/11 with WSL2

• Docker Desktop (running)

• uv (Python package/project manager)

• Claude Desktop

Setup — macOS

1. System tools

xcode-select --install
/bin/bash -c "$(curl -fsSL https://raw.githubusercontent.com/Homebrew/install/HEAD/install.sh)"
echo 'eval "$(/opt/homebrew/bin/brew shellenv)"' >> ~/.zprofile
eval "$(/opt/homebrew/bin/brew shellenv)"
brew install git
curl -LsSf https://astral.sh/uv/install.sh | sh
brew install --cask docker

Open Docker Desktop from Applications and let it finish starting (steady whale icon in the menu bar). Then install Claude Desktop from claude.ai → Download for Mac.

2. Clone and install

git clone https://github.com/vivekpatil200320/cyberrescue.git
cd cyberrescue
uv sync

3. Verify

uv run python -c "from cyberrescue.server import mcp; print('OK:', mcp.name)"
uv run pytest tests/ -v

4. Register with Claude Desktop

Find your uv path:

which uv

Edit (or create) ~/Library/Application Support/Claude/claude_desktop_config.json:

{
  "mcpServers": {
    "cyberrescue": {
      "command": "/Users/YOUR_USERNAME/.local/bin/uv",
      "args": [
        "run",
        "--project",
        "/Users/YOUR_USERNAME/path/to/cyberrescue",
        "python",
        "-m",
        "cyberrescue.server"
      ]
    }
  }
}

If the file already has other mcpServers entries, merge "cyberrescue" in as an additional key rather than overwriting the file.

Fully quit Claude Desktop (Cmd+Q) and reopen it. Check the tools/slider icon near the message box — cyberrescue should appear with all three tools listed.

Setup — Windows (via WSL2)

WSL2 is the recommended path because Docker Desktop for Windows runs its Linux containers through it, and python-on-whales/docker exec behave most predictably there.

1. Install WSL2 and Ubuntu

In an Administrator PowerShell:

wsl --install

Restart if prompted, then open the new "Ubuntu" app from the Start menu and finish the Linux user setup.

2. Install Docker Desktop for Windows

Download from docker.com, install, and during setup enable "Use WSL 2 based engine". In Docker Desktop settings, under Resources → WSL Integration, enable integration with your Ubuntu distro.

3. Inside the WSL Ubuntu terminal — install tooling

sudo apt update
sudo apt install -y git python3 build-essential
curl -LsSf https://astral.sh/uv/install.sh | sh
source ~/.bashrc

4. Clone and install

git clone https://github.com/vivekpatil200320/cyberrescue.git
cd cyberrescue
uv sync

5. Verify

uv run python -c "from cyberrescue.server import mcp; print('OK:', mcp.name)"
uv run pytest tests/ -v
docker ps -a

(docker ps should work inside WSL once Docker Desktop's WSL integration is enabled.)

6. Install Claude Desktop (native Windows)

Download from claude.ai → Download for Windows, install normally (not inside WSL).

7. Register with Claude Desktop

Find your uv path inside WSL:

which uv

Edit %APPDATA%\Claude\claude_desktop_config.json (open via File Explorer: paste %APPDATA%\Claude into the address bar) and add an entry that runs the server through WSL:

{
  "mcpServers": {
    "cyberrescue": {
      "command": "wsl.exe",
      "args": [
        "bash",
        "-c",
        "cd /home/YOUR_LINUX_USERNAME/cyberrescue && /home/YOUR_LINUX_USERNAME/.local/bin/uv run python -m cyberrescue.server"
      ]
    }
  }
}

Replace YOUR_LINUX_USERNAME and the path with your actual WSL username and clone location. Fully quit Claude Desktop and reopen it. Check the tools/slider icon — cyberrescue should appear with all three tools.

Usage

Ask Claude Desktop something like:

Debug the container named my-app: read the last 150 log lines, check its memory and CPU usage, and run printenv DATABASE_URL inside it.

Claude will call the three tools as needed and report back root cause and fix.

Demo containers

demo/ contains three intentionally broken images for testing:

• broken_flask — crashes on startup with a missing-env-var KeyError

• leaking_node — leaks ~10MB/sec until OOM-killed

• crashed_nginx — fails to start due to invalid config syntax

docker build -t demo-broken-flask demo/broken_flask
docker build -t demo-leaking-node demo/leaking_node
docker build -t demo-crashed-nginx demo/crashed_nginx

These same 3 containers, run via infra/docker-compose.yml with fixed names (broken-flask, leaking-node, crashed-nginx), are what the public web demo above is sandboxed to — see src/cyberrescue/demo_policy.py.

Public web demo architecture

• src/cyberrescue/core.py — the actual Docker-calling logic (logs/stats/exec), shared by both paths below.

• src/cyberrescue/server.py — thin @mcp.tool() wrappers around core.py, served over stdio for Claude Desktop. Unchanged behavior from earlier versions.

• src/cyberrescue/demo_policy.py — the allowlist (3 fixed container names) and fixed diagnostic-command menu used only by the public HTTP backend, never by the stdio tools.

• backend/ — a FastAPI app (separate uv workspace member, own dependencies) that wraps core.py under the demo_policy restrictions, adds rate limiting, and an /narrate route that asks Claude to generate a root-cause explanation from already-captured evidence.

• web/ — a Next.js/Tailwind frontend (deployed to Vercel) that talks to the backend.

• infra/ — VPS deployment artifacts (docker-compose for the 3 demo containers, systemd units, Caddy config, a periodic reset job). See infra/README.md to stand it up.

Security

See SECURITY.md for the input validation, command blocklist, and concurrency/sanitization policy.

Future Enhancements

• Standalone binary packaging (PyInstaller/Nuitka) for zero-Python-install distribution

• Streaming log reads for very large logs (currently buffers full log before truncating)

• Optional SQLite audit log for compliance use cases

• Native (non-WSL) Windows support

License

MIT License

Copyright (c) 2026 Vivek Patil

Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the "Software"), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions:

The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software.

THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.