Runtime Guard

Policy enforcement before every tool call. ALLOW, DENY, or ESCALATE in real time.

Three Decisions

ALLOW

Tool call passes policy checks. Execution proceeds normally.

DENY

Tool call violates a policy rule. Execution is blocked and the reason is returned to the agent.

ESCALATE

Tool call needs human approval. Execution pauses until a human approves or rejects.

Python SDK

Import SpiderGuard and wrap your tool execution flow.

python

from spidershield import SpiderGuard, Decision

guard = SpiderGuard(policy="balanced", dlp="redact", audit=True)

# Before tool execution — policy check
result = guard.check("execute_sql", {"query": "DROP TABLE users"})

if result.decision == Decision.DENY:
    print(f"Blocked: {result.reason}")
elif result.decision == Decision.ESCALATE:
    print(f"Needs approval: {result.reason}")
else:
    output = run_tool("execute_sql", {"query": "DROP TABLE users"})
    # After tool execution — DLP scan
    clean = guard.after_check("execute_sql", output)

Policy Rules

Policies are YAML files with pattern-matching rules. SpiderShield ships with three presets: permissive, balanced, and strict.

yaml

# balanced.yaml — ships with SpiderShield
rules:
  - name: block-destructive-sql
    tool: "execute_sql"
    pattern: "DROP|DELETE|TRUNCATE|ALTER"
    match_on: "arguments.query"
    decision: DENY
    reason: "Destructive SQL operations are blocked"

  - name: block-shell-danger
    tool: "run_command"
    pattern: "rm -rf|mkfs|dd if=|:(){ :|:& };:"
    match_on: "arguments.command"
    decision: DENY

  - name: escalate-file-write
    tool: "write_file"
    pattern: ".*"
    match_on: "arguments.path"
    decision: ESCALATE
    reason: "File writes require human approval"

Guard Mode (CLI)

Wrap any stdio-based MCP server with zero code changes.

bash

# Wrap any MCP stdio server — zero code changes
"code-keyword">$ "code-function">spidershield guard \
    "code-keyword">--preset balanced \
    -- "code-function">npx @modelcontextprotocol/server-filesystem /tmp

# With custom policy file
"code-keyword">$ "code-function">spidershield guard \
    "code-keyword">--policy ./my-policy.yaml \
    -- python my_mcp_server.py

# With DLP and audit enabled
"code-keyword">$ "code-function">spidershield guard \
    "code-keyword">--preset strict \
    "code-keyword">--dlp redact \
    "code-keyword">--audit \
    -- "code-function">npx @modelcontextprotocol/server-everything

Proxy Mode (CLI)

Transparent HTTP proxy for network-based MCP servers.

bash

# HTTP proxy mode for network-based MCP servers
"code-keyword">$ "code-function">spidershield proxy \
    "code-keyword">--policy strict \
    "code-keyword">--port 8080 \
    -- python my_http_mcp_server.py

# All tool calls intercepted transparently:
#   read_file("/etc/passwd")  → DENIED
#   exec("rm -rf /")          → DENIED
#   fetch("https://c2.evil")  → DENIED
#   read_file("./data.csv")   → ALLOWED

Policy Presets

Preset	Shell Commands	File Writes	Network	SQL
Permissive	Allow (log only)	Allow	Allow	Allow
Balanced	Escalate	Escalate	Allow	Block destructive
Strict	Deny	Deny	Escalate	Block all mutations

← Documentation DLP Scanner →