MCP Guardian
About
MCP Guardian manages your LLM assistant's access to MCP servers, giving you real-time control over your LLM's activity. It logs all MCP server activity, allows you to approve or deny individual tool call messages in real time, and provides automated safety and privacy scans…
Details
- Author
- eqtylab
- GitHub stars
- 200
- Downloads
- 462
- Categories
- Other, Developer Tools
Jump to
- Message logging – view traces of all LLM MCP server activity
- Real-time message approvals – approve or deny individual tool calls
- Automated message scans – real-time safety and privacy checks (Coming Soon)
- Manage multiple MCP server configurations easily
- Quickly switch between server collections without manual config file editing
Install nix (Linux/macOS) or rustup, nodejs, and yarn (Windows), then clone the repo and run just build-release to produce the mcp-guardian and mcp-guardian-proxy binaries. Use the proxy to intercept and manage MCP server traffic from your LLM host application.
scan_security
Scan MCP server configurations for security vulnerabilities (CVEs, auth, typo-squatting, secrets)
audit_costs
Audit token usage and estimate costs per MCP server
check_health
Check health, latency, and reliability of MCP servers
full_report
Generate a complete security, cost, and health report for all MCP servers
start_behavior_observation
Start observing AI agent tool calls to learn usage patterns for policy generation
stop_behavior_observation
Stop the current observation window and finalize collected data
generate_policy_from_observations
Generate a minimal-privilege YAML policy based on observed tool call patterns
suggest_policy_improvements
Compare observed behavior against current policy and suggest additions/removals
observation_status
Get current behavior observation status and summary
scan_prompt_injection
Scan tool call arguments for prompt injection payloads targeting downstream AI agents
prompt_injection_report
Get prompt injection detection statistics
predict_threats
Generate threat forecast for all configured MCP servers with 30/90/365-day projections
threat_forecast_for_server
Detailed threat forecast for a specific server with risk factors and preemptive hardening recommendations
preemptive_recommendations
Get suggested preemptive policy changes based on threat forecasts
verify_supply_chain
Full supply chain integrity verification with signed attestation for MCP server packages
supply_chain_status
Current trust graph state for all MCP server packages
sbom_export
Export Software Bill of Materials for MCP server packages
detect_drift
Compare current MCP server behavior against a known-good baseline to detect anomalies
capture_baseline
Capture current server state as a known-good behavioral baseline
rollback_server_config
Revert to a previous known-good configuration snapshot
drift_history
List all detected drift events
generate_compliance_evidence
Generate auditor-ready compliance evidence bundle for a framework
compliance_gap_analysis
Identify missing compliance controls and recommend policies
compliance_posture
Get current compliance posture score across all frameworks
list_compliance_frameworks
List all supported compliance frameworks
run_self_assessment
Run a full autonomous red team assessment with attack generation and policy testing
schedule_red_team
Configure periodic autonomous red team assessments
red_team_results
Get latest red team assessment results and recommendations
ab_test_policy
A/B test a proposed policy change against historical attack corpus
contribute_threat_signature
Submit an anonymized threat signature to the cross-deployment intelligence mesh
threat_intel_status
Get mesh connectivity, contribution stats, and known threat feed
deploy_honeypot
Deploy an ephemeral fake MCP server to detect adversarial probing
honeypot_report
Get attack patterns observed by all active honeypots
destroy_honeypot
Tear down a specific honeypot and retrieve captured data
list_honeypots
List all active and destroyed honeypots with summary
negotiate_agent_trust
Initiate an automated trust handshake with another AI agent behind Guardian
agent_trust_status
View all active trust relationships and session details
revoke_agent_trust
Immediately terminate a trust relationship
trust_registry_list
List all registered agents in the trust registry
agentic_status
Get overall status of all agentic AI features including metrics, scheduler, and task queue
compute_trust_score
Compute an A+-F trust score for an MCP server across 8 security dimensions (like SSL Labs for MCP)
scan_response_dlp
Scan MCP tool responses for PII, credentials, sensitive paths, and data exfiltration
certify_server
Run MCP server certification (Bronze/Silver/Gold/Platinum)
list_certified_servers
List MCP servers in the local certification registry with level and expiry
verify_certification
Verify a server certification attestation (JWS) and level
declare_intent
Declare session intent and allowed tools for intent-binding enforcement
run_protocol_fuzzer
Run MCP protocol fuzzer — test defenses against malformed JSON-RPC, overflow, injection
check_sla
Check SLA compliance — p50/p95 latency, error rate, circuit breaker state per tool
run_incident_playbook
Execute an incident response playbook (prompt_injection, credential_leak, shell_injection)
get_agent_reputation
Get agent reputation score — Trusted/Standard/Suspicious/Blocked tier with bypass rate and entropy
harden_config
Analyze MCP server config and get A-F hardening grade with one-click recommendations
detect_collusion
Detect agent-to-agent collusion patterns (recon-then-exploit, coordinated exfil, token sharing)
policy_to_natural_language
Explain MCP Guardian policy YAML in plain English for compliance stakeholders
natural_language_to_policy
Convert a natural-language security goal into a draft YAML policy rule (requires approval before enforce)
query_server_reputation
Query decentralized MCP server reputation (8-dimension consensus score)
quantify_insurance_risk
Compute cyber insurance ALE (Annualized Loss Expectancy) for an MCP server
sample_agent_trust
Thompson Sampling — run Bayesian bandit trust sampling for an agent (Beta posterior, exploration/exploitation)
tune_policy_rule
Contextual Bandit (LinUCB) — select optimal policy action (enforce/relax/skip) based on context
adapt_threshold
SARSA — adaptively tune rate limit, latency limit, or confidence threshold via reinforcement learning
select_fuzz_strategy
REINFORCE — use policy gradient to select optimal fuzzer mutation strategy
MCP Guardian
Overview
MCP Guardian manages your LLM assistant's access to MCP servers, handing you realtime control of your LLM's activity.
📜 Message Logging - See traces for all of an LLM's MCP server activity
💂 Message Approvals - Approve and deny individual tool call messages in real time
🤖 Automated Message Scans - Realtime automated checks for safety, privacy, etc (Coming Soon)
<p align="center">
</p>
MCP Guardian also makes it a breeze to manage multiple MCP server configurations. Quickly switch between server collections without having to manually manage configuration files for your MCP host applications.
Development
This project uses nix to manage a development environment that can be used on Linux and macOS.
Quick Start - Linux / macOS
1. Install nix
2. Enable nix flakes
sudo sh -c 'echo "experimental-features = nix-command flakes" >> /etc/nix/nix.conf'
3. Enter dev shell
nix develop
4. Build project
just build-release
5. mcp-guardian and mcp-guardian-proxy are now in _build/bin/ which is in PATH of the dev shell.
Quick Start - Windows
1. Install git with symlink support (otherwise you will need to restore ./mcp-guardian/bindings before building).
2. Install rustup.
3. Install nodejs.
4. Install yarn.
npm install --global yarn
5. Install .dll dependencies. The easiest way to do this is to install Visual Studio and install the C++ development packages.
6. Install just
cargo install just
7. Enter repo root with git-bash (otherwise just won't work and you'll need to call cargo directly)
8. Build project
just build-release
Justfile
``present just --list``
Available recipes:
build
build-release
clean
do DIR +RECIPE
do-all +RECIPE
fmt
fmt-check
lint
readme-check
readme-update
test
Sign in to leave a review
Use Google, GitHub, or an email account so ratings stay tied to real people.
No reviews posted yet.
