mcp-sandbox

Sandbox any MCP server in one line. Firecracker microVM isolation for Claude Desktop, Cursor, Windsurf, Claude Code, and every MCP client.

Before / After

Before — no sandbox:

{
  "mcpServers": {
    "github": {
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-github"],
      "env": { "GITHUB_PERSONAL_ACCESS_TOKEN": "ghp_..." }
    }
  }
}

Your GitHub token is accessible to the MCP server and its entire dependency tree — 847 transitive npm packages running with full host access.

After — sandboxed in a Firecracker microVM:

{
  "mcpServers": {
    "github": {
      "command": "declaw",
      "args": ["mcp", "--env", "GITHUB_PERSONAL_ACCESS_TOKEN", "--network-allow", "registry.npmjs.org,api.github.com,github.com,codeload.github.com", "--", "npx", "-y", "@modelcontextprotocol/server-github"],
      "env": { "GITHUB_PERSONAL_ACCESS_TOKEN": "ghp_..." }
    }
  }
}

Same MCP server. Same functionality. But now your token can only reach GitHub — even if a dependency is compromised, it can't exfiltrate credentials anywhere else. Only the env vars you explicitly forward with --env reach the sandbox.

Why

MCP servers that connect to external APIs handle your most sensitive credentials — GitHub tokens, Slack bot tokens, API keys, database credentials. These servers run as subprocesses with full host access: your files, your SSH keys, your network.

This isn't theoretical:

• Claude Desktop Extensions had a zero-click RCE rated CVSS 10/10 (LayerX, Feb 2026)
• Cursor had CVE-2025-54135 (CurXecute, CVSS 9.8) and CVE-2025-54136 (MCPoison, CVSS 8.8)

declaw mcp wraps any stdio MCP server in a Firecracker microVM with network deny-all by default. The server works identically — it just can't reach anything you didn't explicitly allow.

When to use this

declaw mcp is designed for MCP servers that talk to external APIs with credentials:

Server	Credentials at risk	Why sandbox it
GitHub	GITHUB_PERSONAL_ACCESS_TOKEN	Token can only reach api.github.com, not exfiltrated elsewhere
Slack	SLACK_BOT_TOKEN	Bot token confined to api.slack.com
Brave Search	BRAVE_API_KEY	API key confined to api.search.brave.com
Database	DATABASE_URL	Connection string can't be sent to external hosts
Any API server	API keys, tokens, secrets	Network allowlist = credential containment

Not a fit for: MCP servers that need local host access (filesystem, SQLite, etc.) — these need your local files to be useful, which a cloud sandbox intentionally prevents.

Install

# macOS / Linux
curl -fsSL https://raw.githubusercontent.com/declaw-ai/declaw-cli/main/install.sh | sh

# or with Go
go install github.com/declaw-ai/declaw-cli/cmd/declaw@latest

# or download binary
# https://github.com/declaw-ai/declaw-cli/releases

Then sign up and authenticate:

# 1. Create a free account at https://console.declaw.ai
# 2. Copy your API key from the dashboard
# 3. Authenticate:
declaw auth login

Client Setup

Claude Desktop

Config path: ~/Library/Application Support/Claude/claude_desktop_config.json (macOS) or %APPDATA%\Claude\claude_desktop_config.json (Windows)

{
  "mcpServers": {
    "github": {
      "command": "declaw",
      "args": ["mcp", "--env", "GITHUB_PERSONAL_ACCESS_TOKEN", "--network-allow", "registry.npmjs.org,api.github.com,github.com,codeload.github.com", "--", "npx", "-y", "@modelcontextprotocol/server-github"],
      "env": { "GITHUB_PERSONAL_ACCESS_TOKEN": "ghp_..." }
    }
  }
}

Cursor

Config path: ~/.cursor/mcp.json — same JSON structure as above.

Windsurf

Config path: ~/.codeium/windsurf/mcp_config.json — same JSON structure as above.

Claude Code

claude mcp add github -e GITHUB_PERSONAL_ACCESS_TOKEN=ghp_... -- declaw mcp --env GITHUB_PERSONAL_ACCESS_TOKEN --network-allow registry.npmjs.org,api.github.com,github.com,codeload.github.com -- npx -y @modelcontextprotocol/server-github

Examples

See examples/ for ready-to-use configs:

• github — GitHub API (repos, issues, PRs, code search)
• brave-search — Web search via Brave Search API
• fetch — Web content fetching and conversion

How it works

declaw mcp is a transparent stdio forwarder. It creates a Firecracker microVM, starts the MCP server inside it, and forwards JSON-RPC messages between the MCP client and the sandboxed server. The client doesn't know anything changed. The server doesn't know it's sandboxed.

Network is deny-all by default. Use --network-allow to open specific hosts the server needs. This is the key security property: credentials passed to the server can only reach hosts you explicitly permit.

Flags

Flag	Default	Description
--network-allow <hosts>	deny-all	Comma-separated outbound hostname allowlist
--template <name>	mcp-server	Sandbox template (default includes Node.js + Python)
--timeout <seconds>	86400	Sandbox timeout (default 24h)
--env KEY or --env KEY=VAL	—	Environment variable to forward (repeatable). KEY reads from host env; KEY=VAL sets explicitly.
--verbose	off	Diagnostic logging to stderr

Custom dependencies

The default mcp-server template includes Node.js and Python, which covers most MCP servers. If your server needs additional system packages (e.g., ffmpeg, native libraries), build a custom template:

# Create a Dockerfile
echo 'FROM declaw/mcp-server:latest
RUN apt-get update && apt-get install -y ffmpeg' > Dockerfile

# Build it (returns a template ID)
declaw template build --dockerfile Dockerfile

# Use the template ID from the build output
declaw mcp --template <template-id> -- your-server-command

See declaw template build --help for details.

Links

• Full CLI documentation
• Declaw docs
• declaw.ai

License

Apache 2.0