CookieCyberTeam
Single Unified Intelligent MCP Server, Defensive Guardrails & Autonomous Cyber Defense System
Next-generation Model Context Protocol (MCP) JSON-RPC 2.0 server enforcing deterministic engineering guardrails, inter-procedural taint flow analysis, air-gapped binary triage, and zero-regression safe patching for autonomous AI coding agents.
Quick Navigation
Overview • The Problem & Solution • System Architecture • Project Genome Profiler • The 5 Guardrail Gates • MCP Interface (16 Tools) • Offline SCA & Binary Triage • CLI & Pre-Commit • Configuration • Verification
Overview
When autonomous LLM agents (Cursor, Claude Code, Windsurf, Antigravity) attempt to fix vulnerabilities or modify production codebases, they frequently introduce regressions, hallucinate non-existent APIs, over-engineer changes, or inadvertently execute destructive shell commands (rm, del, unlink). Furthermore, during security investigations, unconstrained agents can accidentally trigger malware payloads on the host environment.
CookieCyberTeam is a defensive Model Context Protocol (MCP) server engineered to solve these failure modes. Built on pure Python standard library foundations with zero mandatory external runtime dependencies, it provides:
- A Single Unified Intelligent MCP Server: Replaces tool bloat with a dynamic Project Genome Profiler (< 5ms) and an Adaptive Meta-Guide that acts as an LLM Cognitive Anchor.
- 5-Gate Safe Patching Pipeline: Strictly enforces polyglot syntax validation, Ponytail minimalism (stdlib-first, zero dead code), a dual-threshold diff cap ($\le 50$ modified lines), zero-regression inter-procedural SAST, and an absolute zero-deletion invariant.
- Inter-Procedural Call Graph & Taint Engine: Tracks taint flow across intra- and inter-procedural call boundaries across 17 CWE categories with FIRST CVSS v3.1 and v4.0 scoring.
- Air-Gapped Static Binary Triage & Quarantine Vault: Inspects PE, ELF, Mach-O, and DEX headers ($W \oplus X$ violation detection, 1KB Shannon entropy curve, byte-exact IOC offsets) without host execution risk, paired with XOR-encrypted quarantine primitives.
- Multi-Agent DAG Task Scheduler & FIPA-ACL Mailbox: Orchestrates multi-agent incident response pipelines with point-to-point messaging, inbox drainage checks, and automatic orphaned task recovery.
The Problem: Why CookieCyberTeam
| Vulnerability / Failure Mode in Autonomous LLMs | CookieCyberTeam Defensive Engine | Verification & Enforcement Mechanism |
|---|---|---|
| Tool Overload & Context Bloat Installing 5+ MCP servers exposes 70+ tool schemas, burning 15K tokens per turn and causing tool hallucination. |
Project Genome Profiler + Adaptive Meta-Guide Dynamic sub-5ms environment discovery recommends the exact 2–4 tools needed for the user's intent. |
Dynamic JSON genome (mcp://context/project-genome) and cognitive anchor (mcp_adaptive_guide). |
| Guesswork & Hallucinated Fixes LLM modifies source code based on guesses without proving the root cause or validating reproductions. |
Hypothesis-Driven Debugging Workflow Mandates reproducing the failure with a minimal test case before authoring or accepting patches. |
Subprocess/Docker sandbox execution via mcp_execute_sandbox_test and mcp_create_reproduction_test. |
| Scope Creep & Over-Engineering LLM rewrites hundreds of lines, introduces dead code, or adds heavy third-party dependencies for simple tasks. |
Dual Diff Cap & Ponytail Linter (Gate 1.5 & 2) Hard cap of $\le 50$ modified lines ($\le 250$ for new scaffolding), dead code check (YAGNI), and stdlib priority. |
AST analysis and hunk parser in core/guardrails.py reject bloated diffs. |
| Regression Vulnerabilities Fixing an existing bug introduces a new CWE (e.g., SQLi, Command Injection, SSRF, Deserialization). |
Zero-Regression Inter-Procedural SAST (Gate 3) Compares AST findings before and after patching; blocks any new or duplicated CWE findings across 17 categories. |
Inter-procedural call graph visitor (core/ast_scanner.py) + optional Semgrep adapter. |
| Accidental Host Compromise Agent executes an untrusted binary or sample to "inspect its output" or "check dependencies". |
Air-Gapped Zero-Execution Policy Passive structural inspection only. Prohibits launching binary samples on host runtime. |
Structural byte parsing in core/binary_triage.py. Live detonation restricted to external isolated sandboxes. |
| Destructive Actions Agent runs destructive commands ( rm, del, unlink, shutil.rmtree) to clean or organize files. |
Absolute Zero-Deletion Invariant (Gate 4) Destructive file operations are strictly barred. Quarantine uses atomic byte scrambling. |
AST/token scanning rejects destructive primitives. Relocation uses os.replace after XOR encryption. |
System Architecture
┌────────────────────────────────────────────────────────────────────────┐
│ LLM Client (Cursor / Claude Code / Windsurf / AGY) │
└───────────────────────────────────┬────────────────────────────────────┘
│ JSON-RPC 2.0 (stdio / SSE)
▼
┌────────────────────────────────────────────────────────────────────────┐
│ CookieCyberTeam MCP Server │
│ │
│ ┌──────────────────────────────────┐ ┌───────────────────────────────┐ │
│ │ MCP Tools (16 Tools) │ │ MCP Resources (8 Resources) │ │
│ │ • mcp_adaptive_guide (Genome) │ │ • mcp://context/genome │ │
│ │ • mcp_scan_vulnerabilities │ │ • mcp://rules/active-gates │ │
│ │ • mcp_search_code (Semble FTS5) │ │ • mcp://rules/standards │ │
│ │ • mcp_audit_dependencies (SCA) │ │ • mcp://rules/debugging │ │
│ │ • mcp_create_reproduction_test │ │ • mcp://state/agent-context │ │
│ │ • mcp_execute_sandbox_test │ │ • mcp://state/tool-index │ │
│ │ • mcp_preview_surgical_patch │ │ • mcp://playbooks/triage │ │
│ │ • mcp_apply_safe_patch (Roll) │ │ • mcp://playbooks/ir │ │
│ │ • mcp_triage_binary (Air-Gap) ├─┴───────────────────────────────┤ │
│ │ • mcp_run_diagnostic_tool │ MCP Prompts (6 Personas) │ │
│ │ • mcp_submit_dynamic_sandbox │ • Lead Orchestrator │ │
│ │ • mcp_quarantine_artifact │ • Security Auditor │ │
│ │ • mcp_restore_quarantined_file │ • Hypothesis Debugger │ │
│ │ • mcp_generate_containment_rule │ • Patch Developer │ │
│ │ • mcp_terminate_process │ • QA / Code Reviewer │ │
│ │ • mcp_orchestrate_dag (Mailbox) │ • SOC Incident Responder │ │
│ └──────────────────────────────────┘ └───────────────────────────────┘ │
├────────────────────────────────────────────────────────────────────────┤
│ Core Defensive Engines │
│ • Project Genome Profiler: Sub-5ms language, framework & runner probe │
│ • 5-Gate Safe Patch Engine: Hunk JSON + Unified Diff + Auto-Rollback │
│ • Inter-Procedural SAST: Call graph taint propagation for 17 CWEs │
│ • Semble Code Search: Syntactic AST chunking + SQLite FTS5 BM25 │
│ • Supply Chain SCA: Offline OSV vulnerability matcher for pyproject │
│ • Binary Triage: Air-gapped PE/ELF/Mach-O/DEX structural inspector │
│ • Containment Engine: Netsh, iptables, UFW, and XOR-scrambled vault │
│ • DAG Engine: Directed Acyclic Graph runner + FIPA-ACL mailbox │
└────────────────────────────────────────────────────────────────────────┘
Project Genome Profiler & Adaptive Meta-Guide
Instead of burdening the agent's context window with dozens of generic tools, CookieCyberTeam introduces a cognitive anchoring pattern:
[Agent Task Start] ──► [mcp_adaptive_guide] ──► Inspects Repo (< 5ms)
│
┌─────────────────────────────────────────────┴─────────────────────────────────────────────┐
▼ ▼ ▼
[Language & Stack] [Active Guardrails] [Targeted Action]
• Python / TypeScript / Go • Diff Cap: <= 50 lines • Recommended 2-4 tools
• FastAPI / Django / React • Ponytail Linter: Active • Exact test command lines
• Test runner: pytest / unittest • Zero-Deletion: Strictly Enforced • Prohibited destructive ops
Discovery Capabilities (< 5ms, Pure Python stdlib)
- Primary Languages: Python, TypeScript, JavaScript, Go, Rust, Java, C, C++.
- Frameworks: FastAPI, Django, Flask, Express, Next.js, React, Spring Boot.
- Test Runners:
pytest,unittest,jest,cargo test,go test. - Git Context: Current branch, dirty worktree status, protected branch detection (
main,master,prod). - Host Diagnostic Toolchain:
docker,semgrep,r2,ghidra,cfr,jadx,strings,objdump,readelf.
The 5 Patching Guardrails
Every patch processed by mcp_apply_safe_patch or verified via mcp_preview_surgical_patch must pass all five gates sequentially. Failure at any gate triggers an immediate rejection with actionable diagnostic remediation.
graph TD
P[Candidate Patch] --> G1[Gate 1: Polyglot Syntax Dispatch]
G1 -->|Syntax Valid| G15[Gate 1.5: Ponytail Zero-Bloat Linter]
G15 -->|Minimal & Stdlib| G2[Gate 2: Dual-Threshold Diff Cap]
G2 -->|Diff <= 50 / 250| G3[Gate 3: Zero-Regression SAST]
G3 -->|No New Vulnerabilities| G4[Gate 4: Zero-Deletion Invariant]
G4 -->|No Unlink / Rm| G5[Gate 5: Single-Committer Token]
G5 -->|Token Validated| S[Verified Application & Auto-Rollback Guard]
G1 -.->|Syntax Error| R[Reject with Line & Syntax Context]
G15 -.->|Dead Code / Bloat| R
G2 -.->|Diff Cap Exceeded| R
G3 -.->|CWE Introduced| R
G4 -.->|Destructive Action| R
G5 -.->|Unauthorized Committer| R
Gate Breakdown
- Gate 1: Polyglot Syntax Dispatch:
- Python: Strict
ast.parse()validation. - JSON: RFC 8259 validation via
json.loads(). - YAML: Safe YAML parsing with comment and scalar apostrophe tolerance.
- Polyglot (JS/TS/Go/Java/C/C++): Pure-Python comment-aware delimiter stack checking (
(),[],{}) and balanced template string interpolations${...}.
- Python: Strict
- Gate 1.5: Ponytail Zero-Bloat Linter:
- Dead Code Check (YAGNI): Rejects functions/classes with zero call sites in the codebase or tests (unless exported in
__all__). - Stdlib Prioritization: Flags third-party imports when Python standard library equivalents exist, unless already pinned in repository manifests.
- Dead Code Check (YAGNI): Rejects functions/classes with zero call sites in the codebase or tests (unless exported in
- Gate 2: Dual-Threshold Diff Cap:
- $\le 50$ modified lines for edits to existing files.
- $\le 250$ lines for new scaffolding or test fixture generation.
- Gate 3: Zero-Regression Inter-Procedural SAST:
- Scans code before and after patching across 17 CWEs.
- Blocks new CWE types and prevents duplicate finding counts (
patched_counts[cwe] > orig_counts[cwe]). - Call graph visitor discriminates sink types (
sql,command,path,ssrf,deserialization) to prevent cross-procedural false positives.
- Gate 4: Absolute Zero-Deletion Invariant:
- Strictly blocks
os.remove,os.unlink,shutil.rmtree,rmdir, and shell deletion commands (rm,del,Remove-Item). - Quarantined artifacts are XOR-scrambled and atomically moved via
os.replace.
- Strictly blocks
- Gate 5: Single-Committer Isolation:
- Validates ephemeral capability session tokens issued exclusively to
Lead Orchestratorto prevent concurrent write collisions.
- Validates ephemeral capability session tokens issued exclusively to
MCP Interface Reference
Tools (16 Tools)
| Tool Name | Scope & Purpose | Key Parameters |
|---|---|---|
mcp_adaptive_guide |
Dynamic cognitive anchor providing optimal tool sequence and constraints. | task_intent, active_file |
mcp_scan_vulnerabilities |
Inter-procedural AST SAST analysis across 17 CWEs with FIRST CVSS v3.1/v4.0 scoring. | target_path, code_content, delta_only, output_format |
mcp_audit_dependencies |
Offline Software Composition Analysis (SCA) against curated OSV vulnerability database. | path |
mcp_execute_sandbox_test |
Isolated test execution in Subprocess (argv whitelist) or Docker (--network none). |
test_path, sandbox_type, timeout |
mcp_create_reproduction_test |
Enforces Hypothesis-Driven step 1: scaffolds a minimal failing reproduction test case. | save_path, test_code, vulnerability_type |
mcp_preview_surgical_patch |
Dry-run gate verification returning unified diff preview without disk mutation. | target_file, hunks, unified_diff, patched_content |
mcp_apply_safe_patch |
Atomic patch application with dual-format hunk/diff parsing and transactional rollback. | target_file, hunks, unified_diff, committer_token |
mcp_orchestrate_dag |
Multi-Agent DAG workflow management, dependency validation, and FIPA mailbox routing. | action, pipeline_id, task_id, from_agent, to_agent |
mcp_search_code |
Semble-style syntactic AST chunking and SQLite FTS5 BM25 search with disk WAL cache. | query, target_path, mode, top_k |
mcp_triage_binary |
Air-gapped static structural triage of PE/ELF/Mach-O/DEX binaries ($W \oplus X$, entropy, IOCs). | file_path |
mcp_run_diagnostic_tool |
Sandboxed host diagnostic tool execution (strings, readelf, objdump, cfr, jadx, r2). |
tool_name, target_file, args |
mcp_submit_dynamic_sandbox |
Non-blocking submission and polling bridge to external CAPEv2 / Cuckoo sandboxes. | file_path, task_id, async_mode |
mcp_quarantine_artifact |
Atomic XOR scrambling (0x5A) and vault relocation of suspicious artifacts. | file_path, quarantine_dir |
mcp_restore_quarantined_file |
Safely decrypts and restores quarantined artifacts with permission recovery. | quarantine_path, original_destination |
mcp_generate_containment_rule |
Generates cross-platform containment firewall rules (Windows netsh, Linux iptables, UFW, DNS). | target, rule_type, port |
mcp_terminate_process |
Safe process subtree termination (taskkill /T on Windows, process tree signals on POSIX). |
pid, timeout |
Resources (8 Resources)
| Resource URI | Description |
|---|---|
mcp://context/project-genome |
Dynamic JSON genome detailing stack, test runners, git status, tools, and risk profile. |
mcp://rules/active-guardrails |
Live status and operational thresholds for all 5 security guardrail gates. |
mcp://rules/security-standards |
Defensive standards across OWASP Top 10, CWE Top 25, and FIRST CVSS metrics. |
mcp://rules/debugging-mindset |
4-step scientific debugging methodology: Reproduce -> Trace -> Hypothesize -> Confirm. |
mcp://state/agent-context |
SQLite WAL state containing active DAG tasks, dependency graphs, findings, and mailboxes. |
mcp://state/tool-index |
Dynamic catalog of detected host analysis tools (Docker, Semgrep, Radare2, Ghidra, JADX). |
mcp://playbooks/malware-triage |
NIST SP 800-61 Rev 3 static triage and evidence preservation procedures. |
mcp://playbooks/compromise-assessment |
Incident containment, log correlation, and lateral movement detection playbooks. |
Air-Gapped Binary Triage & Quarantine Vault
CookieCyberTeam provides a comprehensive static binary triage suite operating under a strict Zero-Execution Policy:
[Untrusted Binary] ──► [Static Header Parser] ──► Parse Sections (PE / ELF / Mach-O / DEX)
│
├──► Detect W^X Violations (IMAGE_SCN_MEM_EXECUTE & WRITE)
├──► 1KB Windowed Shannon Entropy Curve (Detect Packed/Encrypted Data)
├──► Stream SHA-256 / MD5 (usedforsecurity=False)
└──► Extract IOCs (IPv4, Domain, URL, API Calls) with Exact Hex Offsets
│
▼
[Suspicious Artifact] ──► [Zero-Deletion Vault] ──► XOR-0x5A Scramble ──► Atomic Relocation (.quarantine/)
- PE / ELF / Mach-O / DEX Structural Parsing: Pure-Python unpacking of headers and section tables without external DLL or shared library requirements.
- $W \oplus X$ Memory Protection Violation Check: Identifies sections marked simultaneously executable and writable (
IMAGE_SCN_MEM_EXECUTE | IMAGE_SCN_MEM_WRITE), a hallmark of unpackers, shellcode, and in-memory loaders. - Section Entropy Analysis: Calculates Shannon entropy per section over 1KB sliding blocks to flag packed or encrypted payloads ($\text{Entropy} > 7.2$).
- Byte-Exact Evidence Chain: IOCs (IPv4 addresses, C2 domains, URLs) and suspicious Win32/POSIX API imports are reported with their exact file hex offset (e.g.,
0x0001a4e0). - Zero-Deletion Containment Vault: When quarantining, files are scrambled in-place with XOR byte transformations (0x5A) and moved atomically to
.quarantine/viaos.replace, strictly respecting the zero-deletion invariant.
Standalone Headless CLI & Pre-Commit Hook
CookieCyberTeam provides a standalone CLI (core/cli.py) for automated CI/CD pipelines, local terminal triage, and pre-commit checks:
# 1. Run SAST security scan with exit-code gating
python -m core.cli scan --path . --format json --fail-on high
# 2. Export OASIS SARIF v2.1.0 for GitHub Code Scanning
python -m core.cli scan --path src/ --format sarif > results.sarif
# 3. Supply chain SCA dependency audit
python -m core.cli audit --path .
# 4. Air-gapped static binary triage
python -m core.cli triage --file suspicious_sample.bin
# 5. Generate cross-platform containment firewall rules
python -m core.cli contain --target 198.51.100.42 --rule-type block --port 4444
# 6. Adaptive meta-guide exploration
python -m core.cli guide --intent security_audit
Pre-Commit Integration
Add CookieCyberTeam directly to your .pre-commit-config.yaml:
repos:
- repo: https://github.com/LoveCookieee-java/CookieCyberTeam
rev: v1.0.1
hooks:
- id: cookiecyber-scan
args: ["--fail-on", "high"]
Configuration
CookieCyberTeam automatically discovers .cookiecyber.toml or cookiecyber.json at repository root:
# .cookiecyber.toml
[cookiecyber]
diff_cap_limit = 50
new_file_cap_limit = 250
restricted_branches = ["main", "master", "prod", "production", "release"]
exclude_dirs = ["vendor", "node_modules", ".git", "dist", "build", "__pycache__"]
shannon_entropy_threshold = 7.2
cvss_version = "3.1"
log_level = "INFO"
[cookiecyber.guardrails]
enable_ponytail_linter = true
enable_zero_deletion = true
enforce_single_committer = true
MCP Client Setup
Add CookieCyberTeam to your MCP client configuration (e.g. claude_desktop_config.json, Cursor, Windsurf, Antigravity):
{
"mcpServers": {
"cookie-cyber-team": {
"command": "python",
"args": ["/path/to/CookieCyberTeam/server.py", "--stdio"],
"env": {
"CAPE_API_URL": "https://cape.example.internal",
"CAPE_API_KEY": "optional-api-token"
}
}
}
}
Verification & Benchmarks
CookieCyberTeam is verified by 290 automated unit tests executed across all core defensive engines, achieving a 100% pass rate in under 7.5 seconds.
| Test Suite Module | Target Component Tested | Test Count | Pass Rate |
|---|---|---|---|
tests/test_ast_scanner.py |
Inter-procedural taint flow, 17 CWEs, OASIS SARIF v2.1.0 | 36 | 100% |
tests/test_cvss_calculator.py |
FIRST CVSS v3.1 & v4.0 vector arithmetic & score rounding | 22 | 100% |
tests/test_sandbox_guardrails.py |
5 guardrail gates, polyglot syntax, diff caps, process isolation | 54 | 100% |
tests/test_binary_triage.py |
PE/ELF/Mach-O/DEX parsing, $W \oplus X$ checks, Shannon entropy | 38 | 100% |
tests/test_code_search.py |
Semble syntactic AST chunking, SQLite FTS5 WAL caching | 24 | 100% |
tests/test_sca_scanner.py |
Offline OSV supply chain composition scanner | 18 | 100% |
tests/test_containment_engine.py |
Netsh, iptables, UFW rules, XOR quarantine & restoration | 26 | 100% |
tests/test_project_profiler.py |
Sub-5ms stack discovery, cognitive anchor guide | 20 | 100% |
tests/test_soc_rules.py |
Dynamic MITRE ATT&CK detection engine & playbooks | 24 | 100% |
tests/test_mcp_server.py |
JSON-RPC 2.0 stdio protocol, 16 Tools, 8 Resources, 6 Prompts | 28 | 100% |
| Consolidated Total | Entire CookieCyberTeam Defensive Engine | 290 Tests | 100% Passed |
# Run server diagnostic self-test
python server.py --test-mode
# Run full unittest suite
python -m unittest discover -s tests -v
# Run MCP server over stdio
python server.py --stdio
License
MIT License. See LICENSE for details.
Release files for cookie-cyber-team 1.0.1
For a detailed explanation of source distributions (sdists) and built distributions (wheels), please see the package formats documentation.
Source distribution (sdist)
| File | Size | Uploaded | |
|---|---|---|---|
| cookie_cyber_team-1.0.1.tar.gz | 163.4 kB | Details |
Built distribution (wheel)
| File | Interpreter | ABI | Platform | Reset |
|---|---|---|---|---|
| cookie_cyber_team-1.0.1-py3-none-any.whl | Python 3 | none | any | Details |
Total release size: 285.7 kB