Skip to content
Agentic Control Plane
Security series · Part 3 of 3
MCP Security →

MCP Server Audit Logging: I Checked 8,216 Servers. 74.9% Don't Mention It.

David Crowe David Crowe · · 6 min read
security-research compliance mcp
Share X HN LinkedIn

Your auditor asks: “Show me every action Agent X took on behalf of User Y last Tuesday.”

You check the MCP server logs. There aren’t any. The server doesn’t log tool calls. The agent runtime kept a conversation but not the tool-level details. Your database shows writes, but not which agent or which user started them.

You can’t answer the question. I scanned the MCP ecosystem in March 2026 to see how common that is. It is the normal case.


Methodology

I scanned 8,216 MCP servers across three registries (Official MCP Registry, ToolSDK, Awesome MCP Servers). For each server, I:

  1. Fetched the GitHub README (6,833 came back).
  2. Searched for audit-related terms: audit, logging, trace, tracing, monitor, monitoring, observability, telemetry, access log, activity log, event log.
  3. For servers without a GitHub URL, searched the description and tool metadata for the same terms.

This is a documentation scan, not a code audit. A server could log without saying so. But if your auditor asks “does this system log tool calls?” and the docs are silent, you have a problem anyway. Treat the 25.1% below as an upper bound on documented audit support. The numbers are as of March 30, 2026.

The numbers

Metric Count %
Servers mentioning audit/logging/tracing 2,061 25.1%
Servers with no audit mention 6,155 74.9%
Critical/High sensitivity servers with no audit 1,341 73.0% of critical/high
Servers with any mention of audit logging155467mentions audit: 2,061 (25%)no mention anywhere: 6,155 (75%)a docs-level proxy: a server that never mentions logging is unlikely to give you an audit trail
Three in four of the 8,216 servers never mention audit, logging, or tracing in their documentation. That is a docs-level proxy, but you can't prove what an agent did through a server that logs nothing.

Three out of four MCP servers have no documented audit capability. Servers that handle critical or high-sensitivity data (databases, cloud platforms, payments, email) look the same: 73% have no audit mention.

By category

The categories with the weakest audit coverage:

Category With Audit Total Coverage
Translation services 1 13 7.7%
Support/service management 5 59 8.5%
Art and culture 12 117 10.3%
Location services 14 102 13.7%
Code execution 14 90 15.6%
Customer data platforms 10 56 17.9%
File systems 37 187 19.8%
Finance/fintech 102 470 21.7%
Search/data extraction 98 437 22.4%

Finance has the strictest compliance rules, and only 21.7% of its servers document audit logging. The “monitoring” category does better at 39.6%, which makes sense. But a monitoring tool that logs itself isn’t a record of what your agent did.

The spec gap

The MCP specification (the 2025-12-11 schema) defines transport types (stdio, streamable HTTP, SSE), tool schemas, resources and prompts. It has a logging utility, which lets a server send log messages to the client for debugging.

It does not define:

  • an audit log format
  • a standard way to record tool invocations
  • a way to tie an agent’s action to a user identity
  • a way to ask for the audit trail of one user or session

Server authors have no standard to follow even if they want to. Compare HTTP, where access logs are a convention every web server follows, or databases, where query logging is built in. MCP has no equivalent.

The compliance problem

Agents are going into regulated environments, and the usual frameworks all expect audit trails for systems that touch sensitive data:

Framework Requirement What it means
SOC 2 CC7.2 System activity monitoring You must log and review system activity for anomalies
HIPAA §164.312(b) Audit controls You must record and examine access to ePHI
GDPR Article 30 Records of processing activities You must keep records of data processing operations
PCI DSS 10.1 Audit trail for system components You must link access to system components to individual users

The common thread is that you must tie actions to people. Not to a service account, and not to “the agent.” To the human whose request started it.

MCP servers usually can’t do that. The protocol doesn’t carry the user’s identity, so a server that logs every call still only records “tool X was called with parameters Y.” It can’t say “User Z asked Agent A to call tool X.”

What a useful audit record has

An example shape for one tool call:

{
  "timestamp": "2026-03-30T14:22:01Z",
  "user_id": "user_abc123",
  "agent_session": "session_xyz",
  "tool": "database.query",
  "server": "postgres-mcp",
  "parameters": { "sql": "SELECT * FROM orders WHERE user_id = $1" },
  "result_summary": "47 rows returned",
  "latency_ms": 142,
  "policy_evaluation": {
    "allowed": true,
    "rules_matched": ["allow-read-own-orders"],
    "scopes": ["orders:read"]
  }
}

That record answers who (a named user, not a service account), what (a query against orders), when, whether policy allowed it, and what came back.

The MCP server alone can’t produce any of it. It doesn’t know the user, doesn’t check policy and doesn’t track the session. For 74.9% of servers the docs say nothing about logging.

1,341 high-risk servers with no audit docs

The number that bothers me most: 1,341 servers classed as critical or high sensitivity have no audit documentation. These are database connectors, cloud provisioning tools, payment processors and CRM integrations. They can write to production databases, create cloud infrastructure, send email as a user, read customer records and move money.

For about three in four of them, there is no documented way to answer “what did the agent do, and on whose behalf?”

Where the log can live instead

Asking 8,000 server authors to add audit logging won’t work. Each would pick a different format, storage and retention, and you’d still have no single trail.

The better spot is the place that sees every tool call: the agent’s own tool-call path. That is where a control plane sits. At ACP, a hook in the coding agent checks each tool call before it runs, MCP calls included on the harnesses we cover (the coverage table lists which), and records the user, the tool, the decision and the result in one log. The MCP server doesn’t change.

Two limits. This is a hook on the agent’s machine, so it records what passes through that agent, not what other clients do to the same server. And a workspace starts in audit mode, which records everything and holds only what the safety floor catches. Rules that ask or deny are something you turn on once you’ve seen what’s being called.

For a deeper look, see the reference architecture.


Methodology notes

Audit coverage was assessed by pattern matching on README content and server descriptions, using the terms listed above. This is a documentation scan, not a code audit. A server could implement logging without documenting it, so 25.1% is a ceiling on documented audit capability. That is what a compliance assessor will look at: if you can’t point to documentation showing that the system logs access, “it probably does” won’t pass.

Read the reference architecture → · Get started free →

Running coding agents? ACP records each tool call its hook sees and makes the dangerous shell commands (a force-push, destructive SQL, curl | sh) wait for you. One command, free up to 5 agents:

macOS · Linux · WSL
curl -fsSL https://agenticcontrolplane.com/install.sh | bash
Windows · PowerShell
irm https://agenticcontrolplane.com/install.ps1 | iex

What the installer writes  ·  Getting started

Share X HN LinkedIn
Get the next data drop
What agents actually cost, new tool-surface captures, and the occasional incident post-mortem — sent when we publish something worth your inbox, not on a schedule. Unsubscribe anytime.
Share: Twitter LinkedIn
More in MCP Security
  1. 3. MCP Server Audit Logging: I Checked 8,216 Servers. 74.9% Don't Mention It. · you are here
  2. 5. MCP Gateway Comparison (2026): Composio vs ACP vs DIY
  3. 6. MCP Security Vulnerabilities: 4 Tool Schema Patterns Behind Real CVEs
Related posts

← back to blog