Esquie
MCP server providing computation, encoding, and note-taking tools for AI-assisted reverse engineering, including Python scripting in a sandboxed Docker container.
README
Esquie
MCP server providing computation, encoding, and note-taking tools for AI-assisted reverse engineering. Designed to complement disassembler-specific MCP servers (IDA Pro, Ghidra, Binary Ninja) by handling the ad-hoc computation side of RE work: struct unpacking, address math, crypto checks, encoding/decoding, and arbitrary Python scripting.
Renamed from
re-helper-toolsin 0.3.0. Existing users should remove the old image/container:docker rmi re-helper-sandbox:latest && docker rm -f re-helper-sandbox.
Prerequisites
- Node.js 20 or later
- npm (included with Node.js)
- Docker Desktop or Docker Engine — must be running before using
python_eval
Verify your environment:
node --version # v20.x or later
docker info # should print server info without errors
Quick Start
# Clone and enter the project
git clone <repo-url> && cd esquie
# Install dependencies and compile TypeScript
npm install
npm run build
# Build the Python sandbox Docker image (~1-2 min on first run)
docker build -t esquie-sandbox:latest .
The Docker image is also built automatically on the first
python_evalcall if it doesn't exist, but pre-building avoids a delay during your first session.
MCP Configuration
Claude Code
Add to your project's .mcp.json or ~/.claude.json under mcpServers:
{
"mcpServers": {
"esquie": {
"command": "node",
"args": ["dist/index.js"],
"cwd": "/absolute/path/to/esquie"
}
}
}
cwd must point to the project root so the server can locate the Dockerfile for auto-building the sandbox image.
Claude Desktop
Add to ~/Library/Application Support/Claude/claude_desktop_config.json (macOS) or %APPDATA%\Claude\claude_desktop_config.json (Windows):
{
"mcpServers": {
"esquie": {
"command": "node",
"args": ["/absolute/path/to/esquie/dist/index.js"],
"cwd": "/absolute/path/to/esquie"
}
}
}
Tools Reference
python_eval
Execute arbitrary Python in a sandboxed Docker container. Session state (variables, imports, function definitions) persists across calls within the same server session.
| Parameter | Type | Required | Description |
|---|---|---|---|
code |
string | yes | Python code to execute |
timeout |
number | no | Timeout in ms (default: 30000) |
Pre-installed packages: capstone, lief, pycryptodome, dill. To add packages, build a custom image extending esquie-sandbox:latest and point the server at it via ESQUIE_SANDBOX_IMAGE (see Extending the sandbox). Network is disabled inside the container by design, so pip install from python_eval is not possible.
Examples:
# Expression — result is returned automatically
0x401000 + 0x1a4
# → 4198564
# State persists across calls
from capstone import *
md = Cs(CS_ARCH_X86, CS_MODE_64)
# Subsequent call can use `md`
for insn in md.disasm(b"\x55\x48\x89\xe5", 0x1000):
print(f"0x{insn.address:x}: {insn.mnemonic} {insn.op_str}")
# → 0x1000: push rbp
# → 0x1001: mov rbp, rsp
Hex/Binary Utilities
Native TypeScript tools — no Docker overhead, instant response.
| Tool | Parameters | Description | Example |
|---|---|---|---|
hex_to_dec |
hex |
Hex to decimal (BigInt-safe) | "deadbeef" → "3735928559" |
dec_to_hex |
dec |
Decimal to hex (BigInt-safe) | "3735928559" → "deadbeef" |
hex_to_ascii |
hex |
Hex bytes to UTF-8 text | "48656c6c6f" → "Hello" |
ascii_to_hex |
text |
UTF-8 text to hex bytes | "Hello" → "48656c6c6f" |
xor_buffers |
hex_a, hex_b |
XOR two buffers (shorter repeats) | "4141", "0f0f" → "4e4e" |
hash |
data, algorithm, encoding? |
MD5/SHA1/SHA256 digest | "test", "sha256" → "9f86d08..." |
byte_pattern_search |
hex_data, pattern |
Find byte pattern offsets (?? = wildcard) |
"4d5a900003", "4d5a??90" → {"offsets":[0],"count":1} |
base64_encode |
data, encoding? |
Base64 encode (utf8 or hex input) | "Hello" → "SGVsbG8=" |
base64_decode |
data, output_encoding? |
Base64 decode (utf8 or hex output) | "SGVsbG8=" → "Hello" |
All hex parameters accept optional 0x prefix and ignore whitespace.
Sandbox Management
| Tool | Parameters | Description |
|---|---|---|
reset_sandbox |
(none) | Destroy the container and clear all session state. Next python_eval starts fresh. |
upload_to_sandbox |
filename, content_base64 |
Upload a file (base64-encoded) into /tmp/<filename> inside the container. 10MB limit. |
list_sandbox_files |
(none) | ls -la /tmp inside the container. |
download_from_sandbox |
filename |
Read /tmp/<filename> and return {filename, size, content_base64}. 10MB limit. |
Scratchpad
Key-value store for persisting analysis notes, renamed symbols, struct definitions, and other context. By default in-memory only (cleared on server restart). Set ESQUIE_NOTES_FILE to an absolute file path to persist notes to disk.
| Tool | Parameters | Description |
|---|---|---|
set_note |
key, value |
Store or overwrite a note |
get_note |
key |
Retrieve a note by key |
list_notes |
(none) | List all notes as JSON |
delete_note |
key |
Remove a note |
Notes are also exposed as MCP resources under note://{key} URIs, so MCP clients that support resources can browse and reference them directly.
Sandbox Security Model
The python_eval container runs with multiple layers of isolation:
| Constraint | Effect |
|---|---|
NetworkMode: "none" |
No network access — cannot exfiltrate data or download payloads |
Memory: 512MB |
Hard memory limit prevents runaway allocations |
NanoCpus: 1e9 |
Capped at 1 CPU core |
PidsLimit: 64 |
Prevents fork bombs |
CapDrop: ALL |
All Linux capabilities dropped — zero effective/permitted/inheritable caps |
ShmSize: 1MB |
Shared memory restricted from default 64MB |
ReadonlyRootfs: true |
Filesystem is immutable — only /tmp is writable |
Tmpfs /tmp (100MB) |
Ephemeral writable scratch space, capped at 100MB |
User: sandbox |
Non-root user (uid 1000) inside the container |
no-new-privileges |
Prevents privilege escalation via setuid/setgid binaries |
| Per-call timeout | Default 30s, configurable — kills exec on expiry |
| Output truncation | stdout/stderr capped at 100KB to prevent context flooding |
| Idle auto-expiry | Container destroyed after 30min of inactivity (configurable) |
| Upload/download size cap | 10MB per call to bound exfil-via-roundtrip risk |
| Read-only host mount | When ESQUIE_SANDBOX_MOUNT is set, the host directory is mounted at /mnt/host with the Docker :ro flag — kernel-level read-only. Path is fixed at server start; the LLM cannot select what gets mounted. |
Architecture
Claude Code / Claude Desktop
│
│ stdio (JSON-RPC)
▼
┌─────────────────────────┐
│ MCP Server (Node.js) │
│ │
│ ┌───────────────────┐ │
│ │ hex-utils.ts │──┼── hex_to_dec, xor_buffers, hash, ...
│ │ (native TS) │ │
│ └───────────────────┘ │
│ ┌───────────────────┐ │
│ │ scratchpad.ts │──┼── set_note, get_note, list_notes, ...
│ │ (Map + opt. JSON) │──┼── MCP resources: note://{key}
│ └───────────────────┘ │
│ ┌───────────────────┐ │
│ │ python-eval.ts │──┼── python_eval, reset_sandbox,
│ │ (5 MCP tools) │ │ upload/list/download_from_sandbox
│ └─────────┬─────────┘ │
│ │ calls │
│ ▼ │
│ ┌─────────┴─────────┐ │ ┌───────────────────────────────┐
│ │ sandbox.ts │──┼──────►│ Docker Container │
│ │ (Docker lifecycle)│ │ │ (esquie-sandbox:latest) │
│ └───────────────────┘ │ │ │
│ │ │ python3 /opt/runner.py │
│ │ │ ├─ loads session from pkl │
│ │ │ ├─ exec(code) in namespace │
│ │ │ └─ saves session to pkl │
└─────────────────────────┘ └───────────────────────────────┘
- Lazy init: Container is created on the first
python_evalcall and kept alive for the session. - Session persistence: Python variables survive across calls via
dillserialization to/tmp/session.pklinside the container. - Auto-expiry: Container is automatically destroyed after 30 minutes of idle time (configurable via
ESQUIE_SANDBOX_IDLE_TIMEOUT). - Cleanup: Container is stopped and removed on server shutdown (SIGINT/SIGTERM).
Configuration
Resource limits and timeouts are configured via environment variables:
| Variable | Default | Description |
|---|---|---|
ESQUIE_SANDBOX_MEMORY |
512 |
Memory limit in MB (64–8192) |
ESQUIE_SANDBOX_CPUS |
1 |
CPU core count (1–16) |
ESQUIE_SANDBOX_TIMEOUT |
30000 |
Default exec timeout in ms (1000–600000) |
ESQUIE_SANDBOX_PIDS |
64 |
PID limit (8–1024) |
ESQUIE_SANDBOX_IDLE_TIMEOUT |
1800000 |
Auto-expiry idle timeout in ms (60000–86400000, default 30 min) |
ESQUIE_NOTES_FILE |
(unset) | Absolute path to a JSON file. When set, scratchpad notes persist across server restarts. |
ESQUIE_SANDBOX_MOUNT |
(unset) | Absolute path to a host directory. When set, the directory is bind-mounted read-only at /mnt/host inside the sandbox container so python_eval can analyze its contents without uploading each file. Invalid paths (non-absolute, non-existent, or not a directory) are logged and skipped. |
ESQUIE_SANDBOX_IMAGE |
esquie-sandbox:latest |
Docker image tag the sandbox container is created from. Override to use a custom image (e.g. one that bundles extra Python packages). When set to anything other than the default, the image must already exist locally — the server will not auto-build it. See Extending the sandbox. |
Out-of-range values are clamped to the nearest bound and a warning is logged to stderr.
Set them in your MCP config's env block or export before starting the server:
{
"mcpServers": {
"esquie": {
"command": "node",
"args": ["dist/index.js"],
"cwd": "/absolute/path/to/esquie",
"env": {
"ESQUIE_SANDBOX_MEMORY": "1024",
"ESQUIE_SANDBOX_TIMEOUT": "60000",
"ESQUIE_NOTES_FILE": "/Users/me/.esquie/notes.json",
"ESQUIE_SANDBOX_MOUNT": "/Users/me/samples"
}
}
}
}
Extending the sandbox
The default sandbox image is intentionally minimal: capstone, lief, pycryptodome, dill. The container has no network access by design, so packages cannot be installed at runtime via python_eval. To add tools (e.g. pwntools, unicorn, keystone-engine, yara-python, angr, custom wheels), bake them into a derived image and point Esquie at it.
-
Build the base image once:
docker build -t esquie-sandbox:latest . -
Write a custom Dockerfile that extends it:
FROM esquie-sandbox:latest USER root RUN apt-get update && apt-get install -y --no-install-recommends \ build-essential cmake pkg-config libffi-dev \ && pip install --no-cache-dir --target=/opt/pylibs \ pwntools unicorn keystone-engine yara-python \ && apt-get purge -y build-essential cmake pkg-config \ && apt-get autoremove -y \ && rm -rf /var/lib/apt/lists/* USER sandbox -
Build it:
docker build -t my-esquie-sandbox:latest -f Dockerfile.custom . -
Set
ESQUIE_SANDBOX_IMAGEin your MCP config:"env": { "ESQUIE_SANDBOX_IMAGE": "my-esquie-sandbox:latest" } -
If a previous container exists, force a fresh one so the new image takes effect:
docker rm -f esquie-sandbox
When ESQUIE_SANDBOX_IMAGE is set to a tag other than the default, the server will not auto-build the image — it expects you to have built or pulled it. Missing custom image → first python_eval fails with an actionable error pointing at the build command.
Development
# Run in development mode (auto-compiles via tsx)
npm run dev
# Compile TypeScript to dist/
npm run build
# Rebuild the Docker image (required after changing runner.py or Dockerfile)
docker build -t esquie-sandbox:latest .
# Force-recreate the sandbox container (e.g. after image rebuild)
docker rm -f esquie-sandbox
CI runs npm ci && npm run build on every push and PR to main (.github/workflows/build.yml).
Project Structure
esquie/
├── package.json
├── tsconfig.json
├── Dockerfile # Python sandbox image definition
├── .github/workflows/
│ └── build.yml # CI build check
├── src/
│ ├── index.ts # Entry point: server setup, tool/resource registration, shutdown
│ ├── docker/
│ │ ├── config.ts # Env var config parsing
│ │ ├── sandbox.ts # DockerSandbox class: container lifecycle + exec + file I/O
│ │ └── runner.py # Python runner baked into Docker image
│ └── tools/
│ ├── python-eval.ts # python_eval, reset_sandbox, upload/list/download
│ ├── hex-utils.ts # Native hex/binary/encoding tools
│ └── scratchpad.ts # Key-value notepad (in-memory + optional JSON persistence)
└── dist/ # Compiled output (git-ignored)
Troubleshooting
python_eval fails with "Cannot connect to the Docker daemon"
Docker Desktop or Docker Engine is not running. Start it and try again.
python_eval hangs on first call
The sandbox Docker image is being built automatically. This takes 1-2 minutes on first run. Pre-build with docker build -t esquie-sandbox:latest . to avoid this.
"Conflict. The container name /esquie-sandbox is already in use" A leftover container from a previous session. Remove it:
docker rm -f esquie-sandbox
Session state is lost
The container was destroyed (server restart, Docker restart, manual removal). State lives in /tmp inside the container and does not survive container removal. This is by design.
"Execution timed out"
The default timeout is 30 seconds. Pass a higher timeout value (in ms) for long-running computations. Maximum practical limit depends on the MCP client.
Docker image is stale after editing runner.py
Rebuild the image and remove the old container:
docker build -t esquie-sandbox:latest .
docker rm -f esquie-sandbox
Upgrading from re-helper-tools
Remove the old image and container after upgrading:
docker rmi re-helper-sandbox:latest
docker rm -f re-helper-sandbox
Update any RE_SANDBOX_* env vars in your MCP config to ESQUIE_SANDBOX_*.
Recommended Servers
playwright-mcp
A Model Context Protocol server that enables LLMs to interact with web pages through structured accessibility snapshots without requiring vision models or screenshots.
Magic Component Platform (MCP)
An AI-powered tool that generates modern UI components from natural language descriptions, integrating with popular IDEs to streamline UI development workflow.
Audiense Insights MCP Server
Enables interaction with Audiense Insights accounts via the Model Context Protocol, facilitating the extraction and analysis of marketing insights and audience data including demographics, behavior, and influencer engagement.
VeyraX MCP
Single MCP tool to connect all your favorite tools: Gmail, Calendar and 40 more.
graphlit-mcp-server
The Model Context Protocol (MCP) Server enables integration between MCP clients and the Graphlit service. Ingest anything from Slack to Gmail to podcast feeds, in addition to web crawling, into a Graphlit project - and then retrieve relevant contents from the MCP client.
Kagi MCP Server
An MCP server that integrates Kagi search capabilities with Claude AI, enabling Claude to perform real-time web searches when answering questions that require up-to-date information.
E2B
Using MCP to run code via e2b.
Neon Database
MCP server for interacting with Neon Management API and databases
Exa Search
A Model Context Protocol (MCP) server lets AI assistants like Claude use the Exa AI Search API for web searches. This setup allows AI models to get real-time web information in a safe and controlled way.
Qdrant Server
This repository is an example of how to create a MCP server for Qdrant, a vector search engine.