About Portfolio Cases Services Blog Contact 🎙 Talk to AI
EN DE RU
🎙 Talk to AI
November 22, 2026 · 5 min read

How not to drown in MCP chaos: 50+ remote MCP servers already exist, but without curation your production agent becomes a security hole

I’m Denis Shokhirev, an Agentic AI Systems Architect based in Freiburg im Breisgau, Germany. At DennisCraft AI Studio I ship production AI systems for DACH B2B teams on Claude, Supabase, n8n, Doppler, and self-hosted Postgres, and the MCP problem got real for me the moment an agent tried to use an external tool with broader access than the workflow actually needed. That is the trap with the current MCP rush. More remote MCP servers look like faster progress, but in production they often mean mo

Denis Shokhirev
Denis Shokhirev
Agentic AI Systems Architect
Telegram LinkedIn

I’m Denis Shokhirev, an Agentic AI Systems Architect based in Freiburg im Breisgau, Germany. At DennisCraft AI Studio I ship production AI systems for DACH B2B teams on Claude, Supabase, n8n, Doppler, and self-hosted Postgres, and the MCP problem got real for me the moment an agent tried to use an external tool with broader access than the workflow actually needed.

That is the trap with the current MCP rush. More remote MCP servers look like faster progress, but in production they often mean more hidden trust boundaries, more data paths, and more places where the model can take action you did not intend. A long tool list makes for a good demo. It does not make for a safe system.

MCP standardizes access, not trust

The important distinction is simple: MCP gives models a standard way to discover and call tools, but it does not make those tools equally safe. One server might read customer records, another might execute SQL, another might call internal APIs, and another might write back into operational systems. To the model, these can all look like “available tools.” To your architecture, they are very different risk classes.

Anthropic’s tool-use documentation makes it clear that models can decide when to invoke tools and chain them into a task flow: Anthropic Docs, 2024. That is useful. It also means every added MCP server expands your operational blast radius unless you put strict boundaries around selection, inputs, outputs, and permissions.

The failure pattern I keep seeing

Across recent agent deployments, the same mistakes repeat:

  • tools are granted broader scopes than the business task requires;
  • tool responses return raw records with no redaction or field filtering;
  • tokens, SQL, and customer identifiers leak into logs or traces;
  • write operations are enabled where read-only access would be enough;
  • tool errors are treated by the model as incomplete data, which triggers risky retries or alternate calls.

None of this requires a dramatic exploit chain. It usually starts with convenience.

Why remote MCP servers become production risk so quickly

Every external MCP server is third-party execution logic placed directly on the agent’s decision path. You inherit its authentication model, error handling, output format, logging practices, retry behavior, and release discipline. If you do not review those controls, you are effectively granting your agent access to a black box.

RiskTypical production failurePractical control
Over-scoped accessAgent reads more tables or fields than neededRead-only roles, schema limits, separate service accounts
Secret exposureAPI keys or DSNs appear in stdout, traces, or n8n logsDoppler, output redaction, isolated credentials per tool
Prompt injection via tool outputReturned text contains instructions that conflict with policySchema validation, output sanitization, data/instruction separation
Unsafe write behaviorModel modifies records without approvalHuman review for write/delete, allowlist action paths

OWASP lists prompt injection and excessive agency among the top LLM application risks, and those categories map directly onto MCP-enabled systems: OWASP Top 10 for LLM Applications, 2025. If your model can discover and call tools, then tool curation is part of your security model, not an optional integration detail.

The curation filter I use before any MCP server reaches production

I do not allow an MCP server into production just because it exists or because it has a useful demo. It has to pass a short filter first.

1. Require a narrow contract

If a tool accepts vague natural-language input and returns free-form text, validation is weak and policy enforcement gets blurry. I prefer JSON-shaped inputs and outputs with explicit fields. The less ambiguous the boundary, the easier it is to test and monitor.

2. Split read from write

A tool that both reads and writes business data should be treated as high risk. In most workflows the agent needs visibility first, not mutation rights. Separate those paths. Give the model read-only access by default and move write behavior behind explicit approvals.

3. Put a proxy in front of external tools

I generally do not let a model hit a third-party endpoint directly. I put a small proxy in front of it to validate payloads, enforce limits, and log every call in a structured way.

from pydantic import BaseModel, ValidationError

class LookupRequest(BaseModel):
    customer_id: str
    limit: int

def prepare_payload(payload: dict) -> dict:
    data = LookupRequest(**payload)
    if data.limit > 100:
        raise ValueError("limit too high")
    return {"customer_id": data.customer_id, "limit": data.limit}

try:
    safe_payload = prepare_payload({"customer_id": "c_123", "limit": 20})
except ValidationError as exc:
    print(exc)

4. Isolate credentials per tool

Do not reuse general production secrets across MCP servers. Use separate service accounts, separate secrets, and the smallest scope possible. In my stack that usually means Doppler plus a tightly limited account on the target system.

A practical control pipeline: before launch, at runtime, after incidents

Tool curation is not a one-time checklist. It needs a repeatable control path across the full lifecycle.

bandit -r .
semgrep --config p/owasp-top-ten .
gitleaks detect --source . --no-banner
python -m pytest tests/test_tools.py

Bandit is useful for catching risky Python patterns. Semgrep gives fast coverage for common security anti-patterns. Gitleaks helps prevent secrets from slipping into the repository. None of these tools replace runtime controls, but together they form a solid baseline before a tool is wired into an n8n flow or an agent built on the Anthropic SDK.

Runtime controls matter more than polished demos

Even clean code can behave badly when a model starts selecting tools in live workflows. I recommend three minimum controls for every MCP-enabled agent:

  • an audit log for every tool invocation;
  • rate limits and timeouts on the proxy layer;
  • a policy that blocks write operations unless an approval flag is present.
tool_policy:
  allow_tools:
    - crm_lookup
    - order_status
  deny_write_without_approval: true
  max_calls_per_run: 5
  timeout_seconds: 10
  redact_fields:
    - access_token
    - email
    - iban

Without that, incident response turns into guesswork. You know the output was wrong, but not which tool was called, what payload was sent, or why the model decided to keep going.

From 50+ available MCP servers to the 3–5 worth keeping

Most production agents do not need a marketplace of tools. They need a small set of stable functions tied to a business process. In logistics that may be shipment status, document lookup, and escalation routing. In fintech it may be account status, KYC result retrieval, and transaction summaries. Everything else should wait until observability and failure economics are clear.

My rule is simple: if removing a tool does not break the process SLA, it probably should not be in production yet. “Maybe useful later” is not a valid reason to widen the attack surface. The safest production agent is rarely the one with the most tools. It is the one where every tool is justified, constrained, and observable.

FAQ

Should I avoid remote MCP servers entirely?

No. But you should treat them like high-risk external dependencies, not friendly plug-ins. For many workflows, a narrow internal wrapper is the safer first step.

Is a well-known vendor’s MCP server automatically safer?

Not automatically. Vendor quality helps, but your exposure still depends on scopes, output handling, and how much authority your agent receives inside the workflow.

Is static analysis enough before rollout?

No. Static analysis helps with code and configuration, but it cannot show you how the model will actually choose and chain tools at runtime. You still need runtime logging, limits, and approval gates.

Should agents get write access to Postgres?

Default answer: no. Start read-only, then put writes behind a tightly controlled action endpoint or a human approval path. In most production environments that reduces risk without losing much value.

What can a small team do first if they have no dedicated security engineer?

Start with one proxy layer, one allowlist, one set of checks with Semgrep, Bandit, and Gitleaks, and mandatory audit logging. That already puts you in a much better place than direct model access to every new tool.

If you already run an MCP-enabled agent in production, which stage catches the most issues for you: code review, runtime sandboxing, or business-side incident reporting after the fact? I’d genuinely like to know, because that point usually reveals the real maturity of the system.
I run a free 30-minute stack audit for DACH founders building AI in regulated markets. DM me on LinkedIn or write to @ger_dennis_ai.

For your business

I build agents like this for businesses in Germany and the EU.

Tell me in two sentences what eats your team’s time – reply within 24 hours.

Custom AI agents (German page) →
Continue reading
Your AI agent breaks on mobile UI: how to connect iOS and Android via MCP and run real workflows instead of screenshot demos
One stolen key put 5,000 developers at risk: what the Cursor breach and trojanized Claude Code extension change for AI IDE security
Your customers ask ChatGPT for a tradesman. Does it name your business
Monday morning at a medical practice: what an AI phone assistant takes over and what it does not
All articles →
Where this is applied
Services — what we build
Talk to the voice agent
Case studies
Ready to build?

Turn your process into an AI system

Production quality. DACH B2B focus.

Start a project → ← All articles