Part of the StudioMeyer MCP Stack — Built in Mallorca 🌴 · ⭐ if you use it

mcp-hook-conformance

License Last commit

Test-harness CLI that audits any MCP server for Claude Code v2.1.118 `mcp_tool` lifecycle-hook readiness.

The third foundation pillar:

mcp-protocol-conformance says: does the server speak the protocol?
mcp-server-attestation says: who signed this server?
mcp-hook-conformance says: is it safe to call this tool from a Stop, PreCompact, or UserPromptSubmit hook?

A note from us

We have been building tools and systems for ourselves for the past two years. The fact that this repo is small and has few stars is not because it is new. It is because we only just decided to share what we have built. It is not a fresh experiment, it is a long story with a recent commit.

We love building things and sharing them. We do not love social media tactics, growth hacks, or chasing stars and followers. So this repo is small. The code is real, it gets used, issues get answered. Judge for yourself.

If it helps you, sharing, testing, and feedback help us. If it could be better, an issue is more useful. If you build something with it, tell us at hello@studiomeyer.io. That genuinely makes our day.

From a small studio in Palma de Mallorca.

Install

Terminal

npm install -g mcp-hook-conformance

Or run without install:

Terminal

npx mcp-hook-conformance check ./hook-conformance.config.json

Quick start

bash

# 1. Scaffold a config
npx mcp-hook-conformance init --server-name my-mcp-server

# 2. Edit hook-conformance.config.json (point command/args at your server)

# 3. Run the audit
npx mcp-hook-conformance check ./hook-conformance.config.json

Exit codes:

0 audit passed (no FAILs)
1 audit produced at least one FAIL
2 config error (missing file, invalid schema, unknown suite name)

Subcommands

Subcommand	Args	Description
`check`	`<config-path> [--output=human\|json\|junit] [--suite=...] [--tool=<name>] [--tries=<n>]`	Run audit; emits report to stdout.
`init`	`[--server-name=<name>] [--force]`	Write example `hook-conformance.config.json` to cwd.
`explain`	`<finding-code>`	Print full explanation and remediation for a finding (e.g. `IDEMP-001`).
`version-info`	—	Print tool version and supported MCP-spec range. (Alias: `version`.)

This is a CLI, not an MCP server. Do NOT run mcp-hook-conformance via mcp-inspector — the package declares mcp.specVersion for the targets it audits, not because it speaks MCP itself. Inspector will EACCES on dist/index.js (which is the library export) by design.

Audit suites

Suite	What it checks	Weight
`idempotency`	Same input N times produces same output, no UUIDs/timestamps unless annotated.	30
`side-effects`	Read-only tools must not mutate server state. Requires `stateProbe` config.	25
`latency`	p50 < 30s, p95 < 60s (configurable). 5 samples after 1 warmup call.	20
`determinism`	Response shape stable across calls; no float drift.	15
`dsgvo`	Tool description mentions data flow / retention / deletion (whole-word heuristic).	10

Suite scoring: PASS = 100, WARN = 60, INDETERMINATE = 50, FAIL = 0. Tool score = weighted average across configured suites. Aggregate score = mean of tool scores.

Config format

config.json

{
  "$schema": "https://studiomeyer-io.github.io/mcp-hook-conformance/config.schema.json",
  "server": {
    "name": "my-mcp-server",
    "transport": "stdio",
    "command": "npx",
    "args": ["-y", "my-mcp-server"],
    "env": {}
  },
  "suites": ["idempotency", "latency", "determinism", "side-effects", "dsgvo"],
  "tries": 3,
  "thresholds": {
    "latencyP50Ms": 30000,
    "latencyP95Ms": 60000
  },
  "probes": [
    {
      "name": "search_things",
      "args": { "query": "audit-fixed-input" },
      "stateProbe": {
        "tool": "list_things",
        "args": {}
      }
    }
  ],
  "toolFilter": ["search_things", "get_thing"]
}

probes lets you pin specific arguments per tool and (for read-only tools) a stateProbe so the side-effects suite can detect mutations. Without a stateProbe the side-effects suite returns INDETERMINATE for that tool.

Finding codes

Code	Suite	Severity	Title
IDEMP-001	idempotency	FAIL	Tool output diverges across identical calls
IDEMP-002	idempotency	WARN	Response contains timestamps without `nondeterministic` annotation
IDEMP-003	idempotency	WARN	Response contains UUID/random IDs without `nondeterministic` annotation
LAT-001	latency	FAIL	Tool exceeded p95 latency budget
LAT-002	latency	WARN	Tool p50 latency above 30 seconds
DET-001	determinism	FAIL	Floating-point output drift between calls
DET-002	determinism	WARN	Response shape varies between calls
SIDE-001	side-effects	FAIL	Read-only tool produced state delta
SIDE-002	side-effects	INDETERMINATE	No state-probe configured for side-effect detection
DSGVO-001	dsgvo	FAIL	Destructive tool lacks deletion / retention documentation
DSGVO-002	dsgvo	WARN	Tool description missing data-flow keywords

Run mcp-hook-conformance explain <code> for full remediation text.

Compatibility

Component	Version
MCP spec	2025-06-18
Claude Code	>= 2.1.118 (mcp_tool lifecycle hooks)
Node	>= 20.0.0
`@modelcontextprotocol/sdk`	^1.29.0

Tested transports:

stdio (primary)
HTTP streamable (experimental, audit semantics same as stdio)

Reference smoke tests

The tests/smoke/ suite runs the audit against production tenants of mcp-nex and mcp-crm. Opt-in:

bash

RUN_SMOKE_TESTS=1 npm run test:smoke

Snapshot drift triggers a manual review, not a CI fail.

Limitations (read these)

Side-effect detection is not hermetic. Without a stateProbe we cannot black-box prove a read-only tool is pure. We return INDETERMINATE instead of false-positive FAIL.
Latency is end-to-end stdio roundtrip. Includes JSON-RPC framing, not pure compute.
DSGVO suite is keyword heuristic. It catches missing docs reliably; "looks compliant" is not legal advice.
No auto-fix. This tool reports findings; you fix them.

About StudioMeyer

StudioMeyer is an AI and design studio based in Palma de Mallorca, working with clients worldwide. We build custom websites and AI infrastructure for small and medium businesses. Production stack on Claude Agent SDK, MCP and n8n, with Sentry, Langfuse and LangGraph for observability and an in-house guard layer.