Skip to main content

The MCP firewall for AI agents

Deploy AI agents
without opening
the doors.

Inline security and automatic EU AI Act evidence for regulated teams — one proxy between your agents and every tool and API.

Business trial · 14 days · no credit card · full compliance engine from day one

43%
of public MCP servers allow command injection
33%
allow unrestricted outbound network access
Up to €35M or 7%
maximum EU AI Act fine per non-conformant system, starting 2 Aug 2026

Three public breaches in the last twelve months used MCP as the attack vector — Supabase via Cursor, Postmark, and CVE-2025-6514 in mcp-remote (437,000 downloads). ShieldAgent would have intercepted all three.

Integrations

Works with every tool in your stack

AI Clients

CursorClaude CodeVS CodeGitHub CopilotClineContinue.devWindsurfZedChatGPTGemini CLICursorClaude CodeVS CodeGitHub CopilotClineContinue.devWindsurfZedChatGPTGemini CLI

MCP Servers

SlackGitHubNotionLinearJiraPostgreSQLStripeAWS S3Google DriveCloudflareSupabaseFilesystemSlackGitHubNotionLinearJiraPostgreSQLStripeAWS S3Google DriveCloudflareSupabaseFilesystem

The problem

Nobody is guarding the protocol between AI agents and their tools.

ATTACK SURFACE

Every MCP connection is an open window.

43% of MCP servers contain command injection flaws. 22% expose files outside their intended scope. 5% of open-source MCP servers are pre-seeded with tool poisoning attacks before you install them.

Real incident:

Supabase breach (2025): a Cursor agent read SQL injection payloads embedded in support tickets, exfiltrating the project's service_role key and GitHub OAuth tokens to a public GitHub thread in under 30 seconds.

COMPLIANCE DEADLINE

The EU AI Act clock is running.

August 2, 2026. Fines up to €35M or 7% of global turnover. Eight mandatory sections of technical documentation. Continuous risk management. Automatic event logging. 10-year retention minimum. Over 50% of organizations have no AI inventory.

Cost without tooling:

Enterprise compliance tools cost €50K–€200K/year. A seed-stage startup would burn up to 10% of its entire runway on regulatory paperwork alone.

ACCESS CONTROL

Static RBAC can't keep up with reasoning agents.

Traditional access control was built for deterministic human behavior. An agent that reads emails today may decide to forward them tomorrow — via MCP or a direct HTTP API call — and your permission model has no way to prevent it. Agents have more access than your employees, move faster than your SOC.

The Postmark attack:

The Postmark npm package injected a BCC header into every outgoing email, silently copying attacker@evil.com on password resets, invoices, and internal memos for 14 days before discovery. Static allowlists had no answer.

The insight

Security monitoring data
IS compliance evidence.

If you already sit in the data path — intercepting every tool call, logging every action, enforcing every permission — you already have everything regulators need.

The audit trail? Every tool call you already logged.
The risk management system? Every threat you detected and blocked.
Human oversight evidence? Every escalation you triggered.
Post-market monitoring? The live dashboard you already built.

One proxy, one data path, one platform — and both problems disappear.

<50ms p99

Three pillars

Security and compliance from the same data path.

01

MCP Firewall

MCP + HTTP/REST API — one proxy for every agent connection.

ShieldAgent sits in the data path between your agents and their tools. Every MCP tool call and every HTTP API request passes through before reaching the server. It can stop a threat in 28ms. Observability tools can only tell you what happened — we prevent it.

  • Prompt injection detection (direct + indirect)
  • Tool poisoning & MPMA detection
  • Dynamic least-privilege per agent, per tool, per call
  • HTTP/REST API proxy — same pipeline for Stripe, GitHub, internal APIs
  • Data Loss Prevention — PII, credentials, financial data
  • Human-in-the-loop gates for high-risk actions
See how blocking works
28ms block
02

Compliance Autopilot

EU AI Act documentation that writes itself.

The eight mandatory sections of EU AI Act Annex IV technical documentation are generated from the traffic ShieldAgent already captures. No manual data entry. No consultants. No forms filled out once a quarter.

  • Annex IV all 8 sections — generated from your audit trail, always current
  • Real-time compliance score per AI system (0–100%)
  • Gap analysis with deadline countdown (Aug 2, 2026)
  • ISO 42001 + NIST AI RMF cross-mapping
  • Exportable PDF for auditors
See Annex IV output
94/100
03

Agent Passport

Show, don't tell. Verifiable security per agent.

Stop explaining why your AI agents are secure. The Agent Passport is a shareable, verifiable certificate per agent — a real-time summary of its security posture and compliance status, accessible via public URL. Like SOC 2 trust pages, but for every agent.

  • Shareable URL per agent — show clients instantly
  • Real-time security posture score (0–100)
  • Compliance framework coverage listed
  • Policy controls summary
  • Flywheel: your clients see it, they want it too
Learn more about Agent Passport
grade: A

How it works

From tool call to compliance evidence in under 50ms.

Deploy in shadow mode in 5 minutes. See what would have been blocked. Enable enforcement when ready.

01
Intercept
Every tool call and API request, before it reaches the server

The ShieldAgent proxy receives the agent's MCP tool call or HTTP/REST API request. Zero code changes to your agent — just update the endpoint.

02
Analyze
Policy check, injection scan, DLP in parallel

Policy engine verifies the agent is authorized. Injection scanner checks for malicious payloads. DLP scan ensures no sensitive data escapes.

03
Enforce
Allow or block — in real time, inline

If all checks pass, the request is forwarded. If any check fails, the call is blocked, the agent is notified, and an incident is raised. No downstream damage.

04
Document
Audit trail + compliance evidence, automatically

The full interaction is logged to an immutable audit trail. The compliance engine updates Annex IV documentation and recalculates the agent's risk score.

05
Surface
Dashboard, alerts, and Agent Passport in sync

Security teams see live activity. Compliance officers see their documentation update. Clients see the Agent Passport badge showing verified secure status.

Shadow Mode — zero-risk deployment

Connect ShieldAgent without blocking anything. Get a full report of what it would have blocked, what vulnerabilities it found, and how much your risk score would improve with enforcement active. The same playbook Cloudflare and Datadog used.

--mode=shadow

Agent Passport

Give every AI agent a verified identity.

A cryptographically signed security certificate per agent — risk score, tool permissions, and compliance status in one shareable URL. Issued automatically. No extra configuration.

<!-- ShieldAgent trust badge —— one line -->
<img
src="https://shieldagent.io/badge/apex.svg"
alt="ShieldAgent Certified"
/>
Live — updates automatically·SVG · any size

Risk scoring

Four tiers. No ambiguity.

0–100. Weighted. Updated continuously.

A single number tells your security team exactly how much to trust an agent. Weighted across security events, compliance gaps, integrity checks, and operational health. Score updates every time a new event is logged — recent events weighted more heavily.

Normal0–59

Full throughput — no restrictions.

Elevated60–79

Rate reduced. Review recommended.

High80–89

Heavily rate-limited. Monitoring forced.

Critical90–100

Agent blocked. Manual release required.

Enforcement applied automatically by the ShieldAgent proxy. Full scoring methodology →

Cryptographic certificate

Ed25519-signed passport issued per agent. Tool allowlist, compliance status, and risk score — tamper-evident and independently verifiable.

Verify the signature
CISO procurement view

Share an internal passport URL with your security team before an agent goes live. Risk posture, blocking history, and compliance evidence in one place — no system access needed.

See the internal view
Embeddable badge

One line of HTML drops a live SVG badge into any GitHub README, vendor portal, or marketplace listing. Badge always reflects the current verified status.

Embed the badge

Why existing tools leave a gap

Security vendors don't understand the EU AI Act. Compliance vendors don't understand MCP.

ShieldAgent does both — because they're the same data path. Runtime security + compliance evidence from a single inline proxy.

Langfuse

Observability only — sees what happened, cannot prevent it.

Inline threat blocking (real time)
MCP protocol native understanding
HTTP/REST API proxy (same pipeline)
EU AI Act Annex IV documentation
Runtime access control (per call)
Agent Passport (shareable cert)
Shadow mode deploymentn/a
Compliance from real traffic data
Credo AI

GRC governance — requires manual data entry, no runtime connection.

Inline threat blocking (real time)
MCP protocol native understanding
HTTP/REST API proxy (same pipeline)
EU AI Act Annex IV documentationpartial
Runtime access control (per call)
Agent Passport (shareable cert)
Shadow mode deploymentn/a
Compliance from real traffic datapartial
CyberArk

Machine identity for agents — does not understand MCP protocol semantics.

Inline threat blocking (real time)partial
MCP protocol native understanding
HTTP/REST API proxy (same pipeline)
EU AI Act Annex IV documentation
Runtime access control (per call)partial
Agent Passport (shareable cert)
Shadow mode deployment
Compliance from real traffic data
Promptfoo

Pre-deployment testing — no runtime protection after deploy.

Inline threat blocking (real time)
MCP protocol native understandingpartial
HTTP/REST API proxy (same pipeline)
EU AI Act Annex IV documentation
Runtime access control (per call)
Agent Passport (shareable cert)
Shadow mode deployment
Compliance from real traffic data
Langfuse

Observability only — sees what happened, cannot prevent it.

Credo AI

GRC governance — requires manual data entry, no runtime connection.

CyberArk

Machine identity for agents — does not understand MCP protocol semantics.

Promptfoo

Pre-deployment testing — no runtime protection after deploy.

These tools solve real problems. ShieldAgent complements them by adding the runtime security and compliance layer they don't cover.

Regulatory coverage

Compliance evidence from the data you already capture.

EUAug 2, 2026

EU AI Act

Full Annex IV technical documentation generated from your audit trail. Continuous risk management (Art. 9). Event logging (Art. 12). Human oversight evidence (Art. 14). Post-market monitoring (Art. 72). For Position E (Verdict API) deployments, Annex IV includes the customer's enforcement mechanism and verdict enforcement rates.

  • Annex IV — 8 mandatory technical documentation sections
  • Article 9 — Continuous risk management system
  • Article 12 — Automatic event logging
  • Article 14 — Human oversight mechanisms
  • Article 72 — Post-market monitoring plan
See Annex IV example
ISOAI Management Systems

ISO 42001

AI management system requirements mapped to ShieldAgent controls. Policy framework, risk assessment procedures, and monitoring evidence aligned to the standard.

  • Clause 6 — AI risk assessment & treatment
  • Clause 8 — AI system operations controls
  • Clause 9 — Performance evaluation
  • Clause 10 — Continual improvement
View clause mapping
NISTRisk Management Framework

NIST AI RMF

Govern, Map, Measure, and Manage functions mapped to ShieldAgent's proxy controls, audit trail, and risk scoring system.

  • GOVERN — Policy & accountability structure
  • MAP — AI risk identification & categorization
  • MEASURE — Quantified risk via risk scores
  • MANAGE — Active mitigation via inline blocking
View function mapping
SOC2In progress

SOC 2

Type II — audit in progress

SOC 2 Type II audit in progress. Controls below are live in-product with continuous evidence collection via the Merkle-tree audit trail; third-party attestation report pending auditor sign-off. Position E (Verdict API) deployments are compliant when combined with the customer's own enforcement controls.

  • CC6 — Logical & physical access controls
  • CC7 — System operations monitoring
  • CC8 — Change management evidence
  • A1 — Availability monitoring & SLA tracking
Controls listed are implemented in-product; independent SOC 2 Type II attestation is not yet issued.Request SOC 2 status letter (NDA)

Trust Center

Sub-processors, data handling, incident response SLAs, and responsible disclosure — all in one place.

Open Trust Center
Data residencyEU-first infrastructure · Regional options available
EncryptionTLS 1.3 in transit · AES-256 at rest · Merkle-tree audit logs
RetentionUp to 10 years · configurable per tier
Multi-tenancyComplete data isolation · per-tenant encryption keys

FAQ

Common questions.

EU AI Act · August 2, 2026

Time remaining until EU AI Act enforcement deadline (August 2, 2026): 0 days, 0 hours, 0 minutes, 0 seconds.

Start protecting your agents today.

Book a 20-minute demo, or start the 14-day Business trial and install the proxy in shadow mode in five minutes.

✓ No credit card to start✓ Shadow mode — zero risk✓ EU data residency✓ <5 min setup