MemoryLens MCP

On-demand .NET memory profiling with concrete, AI-actionable code fix suggestions — no profiler to install.

---

Hosted deployment

A hosted deployment is available on Fronteir AI.

Quick Start

npx (any MCP client)

{
  "mcpServers": {
    "memorylens": {
      "type": "stdio",
      "command": "npx",
      "args": ["-y", "memorylens-mcp"]
    }
  }
}

The npm package ships no server code — it is a launcher that installs the MemoryLens.Mcp .NET global tool at a matching version and execs it, so the .NET 10 or 11 SDK must be on PATH. Subsequent starts skip the install entirely and work offline.

VS Code / Visual Studio (via dnx)

Add to your MCP settings (.vscode/mcp.json or VS settings):

{
  "servers": {
    "memorylens": {
      "type": "stdio",
      "command": "dnx",
      "args": ["MemoryLens.Mcp", "--yes"]
    }
  }
}

Claude Code Plugin

/plugin install memorylens --marketplace MarcelRoozekrans/memorylens-mcp

The plugin runs the NuGet tool through dnx, so it needs the .NET 10 or 11 SDK and nothing else.

.NET Global Tool

dotnet tool install -g MemoryLens.Mcp

Docker

docker build -t memorylens-mcp .
docker run -i --rm --pid=host --cap-add=SYS_PTRACE \
  -v /tmp:/tmp \
  -v "$PWD:/workspace" memorylens-mcp

Profiling from a container needs ptrace and the host PID namespace, and on Docker Desktop that namespace is the Linux VM rather than your desktop — see docs/docker.md before choosing this route.

-v /tmp:/tmp is what makes list_processes return anything — the runtime's diagnostic sockets live in the temp directory — and it is also what keeps snapshots alive after --rm, since they are written to /tmp/memorylens-snapshots inside the container. -v "$PWD:/workspace" is only so .memorylens.json is picked up; nothing is written there.

Prerequisites

To run the published tool (dnx, dotnet tool install, the npm launcher):

• The .NET 10 or 11 SDK. The package ships both net10.0 and net11.0 builds.

To build and test from a checkout:

• The .NET 11 SDK, 11.0.1xx feature band: global.json pins 11.0.100 with latestPatch, and SDK 10 alone cannot build the net11.0 target.
• The .NET 10 runtime as well, because the tests run on both net10.0 and net11.0.

Running a filtered subset of the tests, e.g. dotnet test --filter <name>, will exit with code 9 and print error: 1, failed: 0. That's the test project's discovery-collapse guard (--minimum-expected-tests) firing because the filter left fewer tests than expected — it is not a test failure, and a full dotnet test run is unaffected.

How Collection Works

MemoryLens collects heap data in-process over EventPipe, the .NET runtime's built-in diagnostics channel. There is no profiler to install, no download on first use, and no external tool on PATH.

snapshot attaches to a running .NET process by pid, induces a collection, and aggregates the heap into per-type counts and sizes. Snapshots are written as small JSON files under your temp directory and referenced by a short id.

On Linux and in containers, attaching to another process's diagnostic endpoint may require matching UID or SYS_PTRACE — see docs/docker.md.

Available MCP Tools

Tool	Description
list_processes	Lists running .NET processes available for profiling, discovered from their diagnostic IPC endpoints
snapshot	Captures a single memory snapshot of a target process
compare_snapshots	Captures two snapshots with configurable delay and compares them
analyze	Runs the rule engine against a captured snapshot and returns findings
get_rules	Lists all available analysis rules with their metadata

Built-in Rules

ID	Severity	Category	Description
ML001	critical	leak	Event handler leak detected
ML002	critical	leak	Static collection growing unbounded
ML003	high	leak	Disposable object not disposed
ML004	high	fragmentation	Large Object Heap fragmentation
ML005	medium	retention	Object retained longer than expected
ML006	medium	allocation	Excessive allocations in hot path
ML007	medium	retention	Closure retaining unexpected references
ML008	low	allocation	Array/list resizing without capacity hint
ML009	low	pattern	Finalizer without Dispose pattern
ML010	low	pattern	String interning opportunity

Configuration

Create a .memorylens.json file in your project root to customize rule behavior:

{
  "rules": {
    "ML001": { "enabled": true, "severity": "critical" },
    "ML002": { "enabled": true, "severity": "critical" },
    "ML003": { "enabled": true, "severity": "high" },
    "ML004": { "enabled": true, "severity": "high" },
    "ML005": { "enabled": true, "severity": "medium" },
    "ML006": { "enabled": true, "severity": "medium" },
    "ML007": { "enabled": true, "severity": "medium" },
    "ML008": { "enabled": true, "severity": "low" },
    "ML009": { "enabled": true, "severity": "low" },
    "ML010": { "enabled": true, "severity": "low" }
  }
}

Usage Examples

Single Snapshot

Capture a memory snapshot of a running process to inspect current memory state:

> /memorylens
> Take a snapshot of my running API (PID 12345)

Claude will call snapshot with the target PID, then analyze the returned snapshot id and present findings ordered by severity.

Before/After Comparison

Detect memory growth by comparing two snapshots taken with a delay:

> /memorylens
> Check if my app has a memory leak — compare before and after processing 1000 requests

Claude will call compare_snapshots with a delaySeconds value (default 10 seconds) between the two captures, then analyze the diff to identify objects that grew between snapshots.

Family

MemoryLens MCP is part of a small family of .NET projects for AI work:

• AI.Sentinel: Security monitoring middleware for IChatClient: prompt injection, hallucination and PII detection with an audit trail.
• Minos.NET: Typed, source-generated decision questions with calibrated answers.
• Rag.NET: A modular retrieval-augmented generation pipeline on Microsoft.Extensions.AI.
• Roslyn CodeLens MCP: Semantic .NET code intelligence for AI agents over MCP.

Built on ZeroAlloc

MemoryLens MCP uses these ZeroAlloc libraries, which are source generators and analyzers with no reflection at runtime:

• ZeroAlloc.Analyzers

License

MIT