MCP Guardian

A safety layer between your AI assistant and the tools it uses.

Version 4.1.8 · Website · npm · Install & troubleshooting · Changelog

What's new in 4.1.8

• Incident policy workflow — Enterprise AI investigation drawer can generate, preview, accept, or reject blocking rules from incidents
• npm publish hardening — ordered publish script with registry dep resolution checks and clean-install verification before server goes live
• Dashboard stability — fix React hooks ordering in Security/Health panels that caused error #310 on load

What's new in 4.1.7

• Active Rules controls — Security → Policy now includes list/search, soft disable/enable, and hard delete operations synced to YAML
• Policy runtime semantics — enabled: false is honored across rule strategies with backward-compatible defaults
• Policy mutation APIs — cloud + local dashboard endpoints for list/toggle/delete with updated README guidance

What's new in 4.1.6

• mcp-guardian start — one command for proxy + web dashboard on port 4000 (local dev defaults)
• mcp-guardian setup — one-shot install for git clones (pnpm install, build, dashboard SPA)
• npm tarball — prebuilt dashboard UI (deploy/dashboard-spa/out/) built at publish time
• Install guide — expanded troubleshooting in README and docs/INSTALL.md

What's new in 4.1.5

• npm install — fixes broken @mcp-guardian/server@4.1.4 registry manifest (workspace: deps). Use 4.1.5+.

What's new in 4.1.4

• mcp-guardian onboard from global npm — resolves the installed package root (not cwd); writes guardian-configs/ under your current directory; ships scripts/guardian-proxy.sh and policy-audit.yaml in the npm tarball

What's new in 4.1.3

npm install fix — registry manifest now matches published tarballs (^4.1.3 semver deps, not workspace:). Use @mcp-guardian/server@4.1.3 or later. Publish via ./scripts/publish-npm-all.sh (server/CLI ship from .tgz so metadata stays correct).

What's new in 4.1.1

npm install hygiene — fixes supply-chain scanner findings from 4.1.0:

• Published tarballs no longer include postinstall or other lifecycle scripts
• workspace: dependencies are rewritten to semver (^4.1.1) at pack time
• Publish all packages via ./scripts/publish-npm-all.sh (core → plugin-sdk → server → cli)

What's new in 4.1.0

Industry roadmap plan compliance — runtime verification and dashboard wiring for all eleven fleet-wide modules (A1–C5, B1–B3):

• guardian roadmap audit — CLI + GET /api/agentic/plan-compliance/audit verify every shipped module; exit 0 when production-ready
• Dashboard Agentic AI panels — PlanCompliance, Reputation, ZeroTrust, FederatedLearning, Observatory mesh sync, SandboxWizard captured-traffic scorecard, ChainGraph (A1)
• Protection home strip — roadmap compliance score on the main Protection tab with link to Agentic AI
• A1 ONNX graph path — optional fleet chain classifier via GUARDIAN_FLEET_GRAPH_ONNX_MODEL
• B3 MPC-lite masking — pairwise-masked federated gradients (GUARDIAN_FEDERATED_MPC)
• B2/B1 mesh relays — observatory and reputation mesh publish/pull; dev stub via GUARDIAN_OBSERVATORY_STUB
• Docs & env — guardian roadmap * commands documented; production env vars in .env.example

Run guardian roadmap audit --json or open Agentic AI → Overview in the dashboard to confirm 100% compliance.

What's new in 4.0.0

Industry-standard MCP protection — Guardian moves from per-call filtering to fleet-wide, cross-agent security:

• MTX v1 — open threat exchange format (@mcp-guardian/mtx) + cloud hub
• Guardian Certified MCP — HMAC attestation, persistent registry, verification API
• Multi-step attack chains — collusion detector + session-chain graph with proxy enforcement
• Capability graph & intent binding — tool/resource graph and session intent allowlists
• Agent reputation ledger — persistent scores with proxy enforcement
• Dynamic sandbox tiers — shadow / redact / allow with RL-ready persistence
• Protocol fuzzer — expanded corpus with real block validation and cert gates
• Policy simulator — /api/policy/simulate + ab_test_policy MCP tool
• Incident playbooks & AI investigator — webhook/isolate executors; Threat Lab–linked investigations
• Compliance evidence runner — live policy + audit wired to SOC2/HIPAA/PCI/FedRAMP/ISO mappings
• guardian-bench — mcp-guardian bench CLI + public leaderboard

See CHANGELOG.md for 3.4.1 production hardening (JWKS refresh, payload limits, SIEM on all block paths, audit retention).

Roadmap (shipped in 4.0): Semantic policy translator with approval flows, config provenance chain, STRIDE/LINDDUN threat modeling, behavioral biometrics, cross-MCP attack chains with SIEM export, digital twin sandbox, zero-trust SPIFFE scoring, decentralized reputation network, ecosystem observatory, insurance risk quantification + PDF export, and federated threat detection — docs/AGENTIC_ROADMAP.md.

Fleet mandate for CISO buyers

Guardian v4 is designed as a fleet-wide control plane, not a single-proxy filter:

• Mandatory policy provenance — every YAML change is hash-chained, signed, and exportable to SIEM/auditors
• Human-in-the-loop policy approval — NL drafts must pass simulation + explicit approval before apply
• Cross-agent attack chain detection — session graphs span servers; alerts export as CEF for Splunk/Datadog
• SPIFFE/mTLS identity — zero-trust composite scores include workload identity from SPIFFE SVIDs
• Cloud observatory + reputation mesh — anonymized fleet telemetry and server reputation consensus via MCP Guardian Cloud
• Insurance-ready risk reports — ALE quantification with underwriter PDF export for cyber insurance workflows

---

What problem does this solve?

Modern AI assistants (Claude, Cursor, Cline, and others) can connect to tools — read files, run commands, query databases, post to Slack, and more. Those connections often use a standard called MCP (Model Context Protocol).

That power is useful, but risky:

• The AI might read files it should not see.
• It might run shell commands or delete data by mistake or because of a malicious prompt.
• Secrets can leak through tool arguments.
• API costs can spike without you noticing.

MCP Guardian sits in the middle. Every tool request goes through Guardian first. Guardian checks your rules, blocks bad requests, logs what happened, and can show you a live dashboard — before anything reaches your real tools.

Your AI assistant
       │
       ▼
  MCP Guardian  ← reads your rules, blocks bad calls, keeps a log
       │
       ▼
  Your real tools (files, GitHub, database, …)

---

How it works (step by step)

1. You install Guardian and point it at your existing MCP setup (or run mcp-guardian onboard to do this automatically).
2. Guardian wraps your tool servers so the AI talks to Guardian instead of talking to them directly.
3. When the AI tries to use a tool, Guardian receives the request first.
4. Guardian compares the request to your policy (a simple rules file you control).
5. If the request is allowed, Guardian forwards it to the real tool and returns the result.
6. If the request breaks a rule, Guardian blocks it and tells the AI it was denied — the real tool never runs.
7. Every allow and block is saved to a local database so you can review history and see charts on the dashboard.

You stay in control: Guardian does not silently change your rules unless you approve it (for example when reviewing Threat Lab suggestions).

---

Architecture

This section shows how MCP Guardian is wired together: what runs where, how a tool call flows through governance, and how optional Pro pipelines connect to the proxy.

In this section: System overview · Tool call path · Transports · Agentic AI · Dashboard · Pro pipelines · Learning loop

System overview

When you run mcp-guardian start or pnpm dashboard:proxy, one Node process typically hosts the policy proxy, the dashboard API, and (optionally) agentic services. All components share the same audit database (MCP_GUARDIAN_DB_PATH, default ~/.mcp-guardian/history.db).

flowchart TB
  subgraph clients [AI clients]
    Cursor[Cursor / Cline / Claude]
  end

  subgraph guardian [MCP Guardian process]
    Proxy[Proxy layer\nstdio HTTP SSE WS streamable]
  Policy[PolicyEngine\nYAML + hot reload]
  Agentic[Agentic container\noptional hooks]
  DashboardAPI[Dashboard REST + WebSocket]
  end

  subgraph storage [Persistence]
    SQLite[(history.db)]
    SIEM[SIEM exporters\noptional]
  end

  Upstream[Upstream MCP servers\nfilesystem GitHub etc]

  Cursor --> Proxy
  Proxy --> Policy
  Proxy --> Agentic
  Policy --> Proxy
  Agentic --> Proxy
  Proxy --> Upstream
  Upstream --> Proxy
  Proxy --> SQLite
  Proxy --> SIEM
  DashboardAPI --> SQLite
  clients -.-> DashboardAPI

Component	Role	Main code
Proxy layer	Intercepts JSON-RPC; enforces policy on every tools/call	src/proxy/
Policy engine	Evaluates YAML rules, rate limits, RBAC, patterns	src/policy/
History DB	Stores allow/block audit, tokens, cost	src/database/history-db.ts
Dashboard	Local UI + REST API over the same DB	deploy/dashboard-spa/, src/utils/dashboard-server.ts
Agentic	Smart features (injection scan, policy gen, trust, etc.)	src/agentic/

Enterprise deployments may add Redis (rate limits, DPoP, circuit-breaker sync) and PostgreSQL instead of SQLite — see ENTERPRISE_DEPLOYMENT.md.

Tool call path (tools/call)

Every dangerous decision happens before the real MCP server runs. If Guardian blocks a call, the upstream tool never receives it.

sequenceDiagram
  participant Client as AI client
  participant Transport as Proxy transport
  participant PreGuard as Pre-forward guard
  participant Policy as PolicyEngine
  participant Semantic as Semantic gate
  participant Upstream as Upstream MCP
  participant Audit as Audit queue
  participant SIEM as SIEM log

  Client->>Transport: tools/call JSON-RPC
  Transport->>PreGuard: checkExpandedPayload + agentic hooks
  alt blocked at pre-guard
    PreGuard-->>Client: JSON-RPC error -32001
    PreGuard->>Audit: denied record
    PreGuard->>SIEM: tool_blocked
  else allowed
    PreGuard->>Policy: evaluateAsync context
    alt policy block
      Policy-->>Client: JSON-RPC error
      Policy->>Audit: denied record
      Policy->>SIEM: tool_blocked
    else policy pass
      Policy->>Semantic: sync semantic request gate
      alt semantic block
        Semantic-->>Client: JSON-RPC error
        Semantic->>Audit: denied record
        Semantic->>SIEM: tool_blocked
      else forward
        Semantic->>Upstream: forward request
        Upstream-->>Transport: tool result
        Transport->>Transport: response DLP gate
        Transport-->>Client: JSON-RPC result
        Transport->>Audit: allow record
      end
    end
  end

Integration details:

1. Pre-forward guard (src/proxy/tool-call-pre-guard.ts) — caps expanded argument size and runs agentic pre-hooks (prompt injection, etc.) on all transports.
2. Policy (src/policy/policy-engine.ts) — your YAML rules; rate-limit counters survive hot-reload via src/policy/rate-limit-store.ts.
3. Semantic gate (src/proxy/proxy-post-policy-gates.ts) — optional LLM/heuristic check on arguments before forward.
4. Audit — persistCallRecord → async audit-write-queue → SQLite; blocks also emit StructuredLogger.logBlocked for SIEM.

Transports

Guardian implements the same governance stack on every MCP transport your IDE might use:

Transport	Entry module	tools/call governance
stdio	src/proxy/proxy-server.ts	Full pipeline (default for wrapped configs)
HTTP	src/proxy/http-proxy-server.ts	Full + pre-forward guard
SSE	src/proxy/sse-proxy-server.ts	Full + pre-forward guard
WebSocket	src/proxy/websocket-proxy-server.ts	Full + pre-forward guard
Streamable HTTP	src/proxy/streamable-http-proxy-server.ts	Full + pre-forward guard

Run mcp-guardian onboard so client configs point at Guardian-wrapped servers. If an IDE connects to an MCP server around Guardian (common with raw SSE URLs), calls are untracked — metrics and logs will show sse_untracked.

Agentic AI integration

Agentic features are optional modules loaded at boot (src/container.ts). They do not replace your YAML policy; they add observation, scoring, and recommendations.

flowchart TB
  subgraph mcp [MCP surface]
    Tools[MCP tools in src/index.ts]
  end

  subgraph container [DI container]
    Core[agentic/core.ts\npipeline scheduler telemetry]
    Features[Feature modules\npolicy gen injection trust mesh]
  end

  subgraph runtime [Runtime integration]
    Hooks[proxy-integration.ts\npre/post call hooks]
    PreGuard[tool-call-pre-guard.ts]
  end

  subgraph ui [Dashboard]
    API[agentic-dashboard-summary.ts]
    Workspace[Agentic AI workspace SPA]
  end

  DB[(agentic tables\nmigration 011)]

  Tools --> Core
  Core --> Features
  PreGuard --> Hooks
  Hooks --> Features
  API --> DB
  Workspace --> API
  Hooks --> DB

Integration point	What happens
Every tools/call	runAgenticPreForwardHooks can block or sanitize arguments when agentic mode is on
MCP tools	~35 agentic tools registered in src/index.ts for automation and dashboard actions
Modules	40+ agentic modules in src/agentic/ (prediction, policy-gen, mesh, collusion, reputation, etc.)
Dashboard	Agentic AI workspace reads /api/agentic/* summaries
Database	Agentic state in 011-agentic-tables.sql

Module-level detail: docs/AGENTIC_ARCHITECTURE.md · Shipped features: docs/AGENTIC_FEATURES.md · Roadmap: docs/AGENTIC_ROADMAP.md.

Dashboard and observability

flowchart LR
  Proxy[Proxy writes] --> DB[(history.db)]
  DB --> REST[Dashboard REST API]
  REST --> SPA[Next.js SPA\nProtection Activity Agentic]
  REST --> WS[WebSocket push\nGUARDIAN_WS_ENABLED]
  Proxy --> Prom[Prometheus metrics\noptional]
  Proxy --> SIEM[SIEM exporters\nMCP_GUARDIAN_SIEM_ENABLED]

The dashboard is not a separate database — it reads the same call_records the proxy writes. Set MCP_GUARDIAN_DB_PATH consistently when running pnpm real-life:filesystem or other tests so charts match proxy traffic.

Pro pipeline architecture

These Pro workflows run alongside the live proxy. They consume audit data, swarm reports, and LLM output to improve detection — they do not sit in the hot path of every tool call.

Security Swarm

Automated red-team loop: generate attacks, run the harness, detect bypasses, feed learning.

• What it does: Runs scripted steps (build, corpus eval, parity, harness) and records bypasses when policy allows an attack that should be blocked.
• How it connects: Reads/writes under reports/security-swarm/; bypasses and proposals can inform Threat Lab and runtime attack-learning.
• Run: pnpm security-swarm (Pro license in production).

Threat Lab (LLM discovery)

Human-reviewed LLM proposals for new attack fixtures and policy ideas.

• What it does: Collects signals (bypasses, semantic TPs, ThreatIntel), asks a local LLM for new corpus candidates, validates them, writes threat-lab-candidates.json for you to accept.
• How it connects: Outputs feed the adversarial harness and optional policy-applier after review — nothing is applied silently.
• Run: pnpm security-swarm:threat-lab (requires Ollama). See THREAT_LAB.md.

Auto Threat Research

Background LLM research when the proxy blocks suspicious traffic; writes validated adv-*.json fixtures.

• What it does: Debounces block events, classifies attack types, writes harness fixtures when validation passes (dedupe + rate caps).
• How it connects: Uses the same auto-corpus writer as Threat Lab when both GUARDIAN_THREAT_RESEARCH_AUTO and SWARM_THREAT_RESEARCH_AUTO are enabled.
• Run: Enable env flags on the proxy host; or trigger from dashboard Threat Discovery.

Continuous improvement loop

flowchart LR
  Live[Live proxy blocks] --> Audit[(history.db)]
  Live --> Learn[Attack learning]
  Swarm[Security Swarm] --> Bypasses[bypasses.json]
  Bypasses --> ThreatLab[Threat Lab]
  ThreatLab --> Harness[adversarial harness]
  AutoResearch[Auto Threat Research] --> Fixtures[adv fixtures]
  Fixtures --> Harness
  Harness --> Policy[Policy YAML updates]
  Policy --> Live

Deep dive: docs/ARCHITECTURE.md.

---

Features explained

Below is what each major capability does, in plain language.

Policy proxy (the core)

What it is: A filter on every tool call.

How it works: You write rules in a YAML file (see The policy file below). Rules can allow specific tools, deny dangerous ones, limit how often tools run, cap token usage, and match patterns in arguments (for example “block if the path contains ../”). When you change the file, Guardian can reload rules without restarting.

Why it matters: This is your main line of defense — fast, predictable, and fully under your control.

---

Attack blocking (built into the default policy)

What it is: Hundreds of pre-written checks for common abuse.

How it works: Before a call reaches your server, Guardian looks for things like shell commands hidden in arguments, path traversal (../etc/passwd), SQL injection patterns, attempts to exfiltrate secrets, suspicious URLs, and Unicode tricks that hide malicious text. If a pattern matches, the call is blocked and logged.

Why it matters: Many real-world attacks look like normal tool calls; these checks catch a large class of them without an AI model.

---

Cost tracking

What it is: A running tally of how much your tool usage costs.

How it works: Guardian estimates tokens and dollar cost per call (using model pricing when available). You can set budgets and see burn rate over time in the dashboard.

Why it matters: Runaway agents or loops can get expensive; you see it early.

---

Health monitoring

What it is: A health check for each connected MCP server.

How it works: Guardian tracks success rate, latency, and whether a server is responding. If a server keeps failing, a circuit breaker can stop hammering it.

Why it matters: You notice broken or flaky integrations before users complain.

---

Live audit log

What it is: A permanent record of what was allowed and what was blocked.

How it works: Each decision is stored in a local SQLite database (default: ~/.mcp-guardian/history.db). The dashboard reads this database to show tables, charts, and filters.

Why it matters: Security and debugging need a clear trail — who tried what, when, and why it was blocked.

---

Package scanning (CVE and typo-squat)

What it is: A check on MCP packages before you trust them.

How it works: Guardian can scan installed or configured packages for known security issues (CVEs) and names that look like famous packages but are slightly misspelled (typo-squatting).

Why it matters: Supply-chain attacks often arrive as “almost the right” package name.

---

Adversarial harness (offline tests)

What it is: A large automated test suite that fires attack-like requests at your policy without a live AI.

How it works: Run pnpm harness from the repo. It replays 800+ fixtures and reports what would be blocked or allowed.

Why it matters: You can change rules and immediately see if you broke legitimate use or left a hole open.

---

Real-life scenarios (live tests)

What it is: A short or long run of real attack traffic against a real filesystem MCP server through Guardian.

How it works: Commands like pnpm real-life:filesystem drive the official filesystem server with path traversal, injection, and similar tests while the proxy is running. Results show up in the dashboard if you use the same database path.

Why it matters: Offline tests are fast; live tests prove the full path (proxy → policy → log → UI) works.

---

Agentic AI features (version 4.1)

These are smart assistants inside Guardian that watch, score, and recommend — they do not replace your policy unless you choose to apply a suggestion.

Shipped today