uta-shield · page 01 / 08
AliceLabs LLC · UTA stack · v0.2.2 shipped

UTA
SHIELD

The wall for the other side of MCP.

Lasso, Mintive, Obot and agentgateway protect the consumer — the company whose agents call third-party servers. Nobody protects the operator: the SaaS that publishes an MCP server and exposes its database to every anonymous agent on the internet. uta-shield is the mirror side. Wrap your server once — identity gates, threat-intel blocking, rate limits, DLP, x402-ready metering and a full audit trail. Zero code changes. Zero dependencies.

Lasso, Mintive, Obot y agentgateway protegen al consumidor — la empresa cuyos agentes llaman a servidores de terceros. Nadie protege al operador: el SaaS que publica un servidor MCP y expone su base de datos a cada agente anónimo de internet. uta-shield es el lado espejo. Envuelve tu servidor una vez — gates de identidad, bloqueo de threat-intel, rate limits, DLP, metering listo para x402 y auditoría completa. Cero cambios de código. Cero dependencias.

85/85 tests· 0 dependencies· 58 kB· 23 live IOCs· ~12 h takedown· 2 min install

npx uta-shield wrap --upstream http://localhost:3001 --port 8787

page 02 / 08 — the gap

The gap

one category, one empty column

Roughly half of organizations already experiment with MCP; only ~11% run it in production. Thousands of MCP servers are publicly reachable — several hundred have leaked credentials. A funded security category grew out of this: all of it on the consumer side.

The operator who publishes the server — the one actually holding the database — has nothing. That empty column is this product.

Cerca de la mitad de las organizaciones ya experimenta con MCP; solo ~11% llega a producción. Miles de servidores MCP son alcanzables públicamente — varios cientos han filtrado credenciales. De ahí creció una categoría de seguridad financiada: toda del lado consumidor.

El operador que publica el servidor — el que realmente sostiene la base de datos — no tiene nada. Esa columna vacía es este producto.

consumer side — covered
“My employees' agents must not touch dangerous servers.”
Solved by Lasso, Mintive, Obot, Solo.io agentgateway, Cequence. Crowded, funded, well-marketed.
operator side — this
“I publish an MCP server. Every anonymous agent on the internet hits my database.”
Solved by uta-shield — the first provider-side firewall: 58 kB, auditable line by line, shipped.
why nowx402 already settles tens of millions of agent-paid calls a month. The faster agents pay, the faster operators need a wall.
positionNot another gateway. The gate that faces the server.
page 03 / 08 — eight layers

What it does

solid underline = shipped · dashed = design hook — honesty policy
  1. Agent identity — x-agent-name / passport header, optional /.well-known/agent.json verification, --require-identity refuses anonymous calls.shipped
  2. Reputation — per-agent score built from identity verdicts and behavior.hook
  3. Threat intel in/out — IOC matching on domains, URLs, IPs, emails, hashes, wildcards, mutexes. Applied to inbound args and outbound responses.shipped
  4. Rate limits — token bucket per agent+tool, free/paid tiers. Abuse gets 429 before it gets your database.shipped
  5. Billing — per-call metering + x402-ready pricing manifest at GET /pricing; settlement plugs into your payment layer.metering
  6. DLP — emails, JWTs, AWS keys, API tokens, private keys, cards masked in responses. Your API can't leak what it never emits.shipped
  7. Reverse drift — notify clients when your schema changes. Fewer broken agents, more trust.hook
  8. Audit — JSONL per decision. Indicator + campaign recorded for triage, never raw arguments. SIEM-grade hygiene.shipped

~1,300 auditable lines of Node ≥18 (ESM), hand-rolled MCP stdio (newline-delimited JSON-RPC 2.0, protocol 2025-06-18). A security product with 47 transitive dependencies is an attack surface, not a defense.

page 04 / 08 — use case

Day one on a CRM server

the same data class Apollo.io exposes via mcp.apollo.io

You publish an MCP server that exposes contacts, pipelines and email history. Today any anonymous agent can hammer it, exfiltrate PII through it, or use it as a lookup oracle. You wrap it once:

npx uta-shield wrap --upstream http://localhost:3001 --port 8787

From then on agents talk to :8787. Four things happen on day one — the four verdicts below.

Publicas un servidor MCP que expone contactos, pipelines e historial de email. Hoy cualquier agente anónimo puede golpearlo, extraer PII a través de él, o usarlo como oráculo de consulta. Lo envuelves una vez:

npx uta-shield wrap --upstream http://localhost:3001 --port 8787

Desde ahí los agentes hablan con :8787. El día uno pasan cuatro cosas — los cuatro veredictos de abajo.

verdict 01
ALLOW
A legitimate client calls search_contacts. Call passes, 23 ms, metered to the ledger.
identity=copilot-cli · tier=free → ledger.jsonl
verdict 02
DLP
The response contains the fetched record. Two emails masked, one AWS key blocked — the wall emits it clean.
2 masked · 1 blocked → response leaves clean
verdict 03
BLOCK_INTEL
An argument references gaganata.ink — a live C2 from the bundled feed. Blocked, with evidence attached so the agent can reason about WHY.
domain=gaganata.ink · role=C2 · campaign=taikotalent-2026-10
verdict 04
429
Same agent retries forty times a minute. The bucket empties; your upstream never sees it.
rate_limit · agent=anonymous · tool=search_contacts
$ curl -s -H "x-uta-admin: $UTA_ADMIN" :8787/audit | tail -2 {"ts":"…","seq":12,"verdict":"BLOCK_INTEL","tool":"fetch_url","reason":"IOC match in tool-call arguments","findings":[{"value":"gaganata.ink","campaign":"taikotalent-2026-10"}]} {"ts":"…","seq":13,"verdict":"ALLOW_VERIFIED","agent":"copilot-cli","tool":"search_contacts","latency_ms":23}
Every decision leaves a line — operator telemetry only. No admin token configured → /audit answers 404 (fail-closed); wrong token → 401, and the denial is audited. Triage reads verdicts and campaigns — never your users' raw arguments.
page 05 / 08 — live intel

Not promises — a real campaign

campaign taikotalent-2026-10

Most security products launch with an empty database. uta-shield ships with a documented, operationally-processed phishing campaign: fake-recruiter malware against open-source developers, delivered through “terminal-as-CAPTCHA” social engineering.

AliceLabs ran the takedown — the phishing domain fell in ~12 hours after coordinated reports to 13 channels (AWS abuse case 179121234700301, registrar ticket NC-DFA-2100). The C2 infrastructure is still tracked live. That whole case is the bundled feed.

La mayoría de los productos de seguridad arranca con base vacía. uta-shield viene con una campaña de phishing documentada y procesada operativamente: malware fake-recruiter contra desarrolladores open-source, entregado con ingeniería social “terminal-as-CAPTCHA”.

AliceLabs operó el takedown — el dominio de phishing cayó en ~12 horas tras reportes coordinados a 13 canales (caso abuso AWS 179121234700301, ticket registrar NC-DFA-2100). La infraestructura C2 sigue rastreada en vivo. Todo ese caso es el feed incluido.

agents, toouta-shield mcp runs the intel as tools: lookup_ioc, check_url, screen_text — every response is DATA, nothing is fetched or executed.
page 06 / 08 — the project behind

The project behind

AliceLabs LLC — Wyoming, USA

uta-shield is not a weekend script — it is the security layer of the UTA stack: the conformance tooling, the agent trust credentials and the agent-first marketplace infrastructure that today serves 71,000+ audited skills. The wall is the piece that lets agents and servers do business without trusting each other blindly. Everything below is live, installable and versioned today.

uta-shield no es un script de fin de semana — es la capa de seguridad del stack UTA: las herramientas de conformance, las credenciales de confianza de agentes y la infraestructura de marketplace agent-first que hoy sirve más de 71,000 skills auditadas. El muro es la pieza que permite que agentes y servidores hagan negocios sin confiar a ciegas. Todo lo de abajo está vivo, instalable y versionado hoy.

  1. UTA — Universal Trust Adapter. Conformance suite for agents and servers: schema, interop, certification vectors. @marketnow/[email protected] on npm.npm
  2. Alethech. Verifiable agent memory — append-only, evidence-backed, portable across compatible agents. alethech on PyPI.pypi
  3. MarketNow. Agent-first marketplace: 71,198 skills audited, agent-hire, m2m-checkout, Agent Trust Credentials — under MAP v1.1.0.live
  4. uta-shield. The provider-side firewall + PhishGuard threat-intel MCP — the trust enforcement layer of the stack.this
page 07 / 08 — license

Licensing

licensed for use — run it free, pay when it becomes your product

The package carries a license for use (see the LICENSE file): free to run on your own infrastructure, including commercially. Not free to redistribute, resell, or offer as your own hosted service. Commercial tiers add managed intelligence and hands-on deployment. USD or USDC.

El paquete trae una licencia de uso (ver el archivo LICENSE): gratis para correrlo en tu infraestructura, incluso comercialmente. No gratis para redistribuirlo, revenderlo ni ofrecerlo como servicio hospedado propio. Los tiers comerciales agregan inteligencia gestionada y despliegue asistido. USD o USDC.

self-host
Free
license for use, forever
  • All shipped layers — identity, intel, limits, DLP, metering, audit
  • Bundled campaign feed (taikotalent-2026-10)
  • Your infrastructure, your data
S7 — managed feed
$50/mes
per organization
  • Updated campaign feed + webhook alerts
  • YARA rules for payload families
  • Free for the first 100 ATC-holding agents
S6 — assessment + pilot
$1,000 + $100/mes/server
hands-on with the AliceLabs team
  • Security audit of your MCP server
  • Deployed pilot — identity + intel + audit live
  • Written report + hardening roadmap

Commercial exceptions (white label, OEM, redistribution, managed service, feed resale): [email protected]