No perimeter
AI agents read files, push code, and query databases autonomously — with no enforcement layer between the agent and your infrastructure.
MCP security platformOpen source
Perimeter security
for your AI.
Runtime enforcement, policy control, and trust scores for every MCP tool call. mastyf.ai intercepts agent actions, blocks violations before they execute, and scores npm packages so teams ship without guessing.
- Runtime enforcementBlockGuard intercepts every tools/call before it reaches your infrastructure — fail-closed by default.
- Package trust scoresInstant 0–100 scores and public badges for any npm MCP package — no account required.
- Cloud control planePolicy, fleet, keys, and evidence from one console — open source and self-hostable.
Runtime enforcement
BLOCKED
filesystem/read
Path traversal · /etc/passwd
github/create_prshell/exec
Try it now — no account required
97.50%
Hybrid coverage
75% deterministic + 95% neural · 3,000 enterprise
0.00%
False positives
Wilson 95% [0.00,0.19%] · 2,000 benign
17.3µs
Fast path
P50 · 1.1GB CPU · zero GPU
10.5281
Zenodo DOI
Research preprint v2.0 · CC-BY 4.0
One platform. Every layer of MCP security.
Runtime enforcement, policy control, ops visibility, and public trust scores — built from the open-source mastyf.ai repo.
Every tool call inspected before it runs
mastyf.ai sits between your AI client and MCP servers. BlockGuard enforces policy synchronously on every call — prompt injection, path traversal, secret exfiltration, and shell commands are stopped before they reach your infrastructure.
- Three-layer detection: regex, schema, semantic LLM
- Fail-closed — blocked calls never execute
- Full audit trail with allow/block status
localhost:4000
Live tool-call feed
- BLOCKEDfilesystem/readblock-sensitive-paths
/etc/passwd
BlockGuard enforced blocked on filesystem/read
Why mastyf
Built for agents that act — not just chat
Generic AI firewalls describe risk. mastyf enforces on every MCP tool call with repo-backed policy, corpus gates, and a Security Swarm that compounds with every attack.
No audit trail
When something goes wrong, teams cannot answer what action the agent took, on behalf of which user, in which system.
No trust signal
Teams have no simple way to verify which MCP packages are safe before agents connect to production data.
Enforce on every call
BlockGuard sits in the MCP path and stops violations before execution. Pattern detection runs in microseconds; semantic LLM audit catches borderline cases async.
- 1
Deploy the proxy
Run mastyf.ai between your AI client and MCP servers — Docker, source build, or connect to the cloud console. Every tool call flows through BlockGuard.
- 2
Define your policy
Start in audit mode to see what your agents do. Tune rules in YAML, then switch to block mode for production enforcement.
- 3
Score, badge, and ship
Look up npm MCP packages for trust scores, embed badges in READMEs, and let the Security Swarm keep learning from every block.
Threats stopped at runtime
- Prompt injection
- Path traversal
- Secret exfiltration
- Shell injection
- Data exfiltration
- SSRF
- Encoding evasion
- Cost abuse
- Rug-pull attacks
Interactive architecture
Holistic MCP protection across ingress, economics, policy, intelligence, upstream, and egress. Every tools/call on all transports flows through the defense orchestrator.
View static diagram
Trust, compliance, and deployment proof
Real capabilities from the open-source repo — fleet management, evidence packs, economics controls, and human-reviewed threat discovery.
Multi-server MCP protection at scale
Auto-discover MCP servers across your org, patch IDE configs, and enforce policy fleet-wide from the cloud console. One control plane for every agent endpoint.
- Auto-discovery of MCP endpoints
- IDE config patching (Cursor, Claude Desktop)
- Centralized policy rollout
228/228Corpus gates · 0 bypasses
6-phaseDefense Fabric on every call
HelmEnterprise & throughput profiles
OWASPAttack matrix evidence mapping
Research preprint · Zenodo DOI 10.5281/zenodo.22206627 · CC-BY 4.0
The security your agents were missing
Give your AI tools a real perimeter. Every call is checked against what it was actually allowed to do — so hijacks fail, data stays inside, and your team ships faster.
Why this matters to you
Your AI agent reads docs, calls tools, and moves money — but it cannot tell a trusted instruction from attacker data hidden in a webpage or tool output. Mastyf Guard puts a perimeter around every tool call so bad instructions never execute, and good work never gets blocked.
- Stops hijacked tools coldEven if the model is tricked, out-of-scope tools cannot run — zero chance.
- Catches data theft across toolsTracks data across steps and blocks leaks before they leave — 1,152 cases proven.
What you get — proven at scale
99.33%Attacks blocked · tested on 50,000 real cases · instant
97.50%Coverage where it counts · enterprise tasks stay safe
0.00%False alarms in production · 2,000 real team workflows
17.3µsFeels instant · runs on a tiny 1.1GB CPU, no GPU needed
Four simple guarantees: right destination, right scope, no privilege creep, spend stays capped. Routine work flows with 0% extra AI cost.
Das, Rudraneel. “Capability-Mediated Perimeters for Secure AI Agent Tool Execution.” Zenodo Research Monograph, v2.0, 2026. doi: 10.5281/zenodo.22206627 · CC-BY 4.0 · mastyf-guard.pdf
From research preview to production
The 1.5B guard is gated on Hugging Face for research and enterprise preview. Commercial production requires an active Mastyf Enterprise Pass.
Research & Preview
Try the frozen v2.0 baseline locally.
- Gated HF repo — share contact to access weights
- Ollama, vLLM, llama.cpp, Docker — Q4_K_M 986 MB
- Paper + 9 figures + 28,450 train / 3,550 val
Most teams
Enterprise Pass
Commercial production deployments.
- License key for production tool dispatch
- Fleet policy, evidence packs, support
- Cloud console — policy, keys, fleet, audit
Open source
Perimeter proxy, swarm, and trust scores.
- AGPL 3.0 — self-hostable
- 228/228 corpus gates · Security Swarm
- github.com/mastyf-ai/mastyf.ai
Cloud console
Sign in with Google or GitHub to edit policy YAML, copy tenant env snippets, rotate API keys, and manage your fleet. Free — no credit card.
Your questions, answered
Runtime enforcement, trust scores, Security Swarm, and deployment.
mastyf.ai is perimeter security for AI agents using MCP. It intercepts every tool call, enforces your security policy, blocks violations before execution, and provides trust scores for npm MCP packages — all from one open-source platform.
Contact us
Questions about scores, badges, the cloud console, or privacy requests?
Email mastyf.support@gmail.comReady to score your MCP servers?
Look up any npm package free — no account required.