Skip to main content

agent-slowdown

Your Mac or PC crawls when Claude Code, Codex, Cursor and Antigravity run side by side. This finds out why, and fixes the parts that can be fixed.

top makes this hard to see. It lists one row per process, so forty 90 MB copies of the same helper rarely reach the first screen. This tool sums by name.

pipx run --spec git+https://github.com/tonydzi/agent-slowdown agent-slowdown

Python 3.9+, no dependencies, macOS, Linux and Windows. That command is read-only: nothing is stopped or changed unless you pass --kill.

== Biggest groups ==
  memory  copies  cpu%  name
  3.8 GB      10    15  claude
  2.3 GB      31     4  ChatGPT

== Orphans ==
Orphaned agent helpers (parent gone, older than 60 min): 3, 3 MB
  29462    1 MB     0  82.9h  /bin/zsh -c source ~/.claude/shell-snapshots/snapshot-zsh-1...

Those three were real, found on our own machine: wait loops left by dead sessions, still polling every 15 seconds for a file that would never appear, 88 hours later.

Six causes, what helps, and how sure we are

# Cause Command What helps Status
1 A stdio MCP server starts a copy in every session agent-slowdown mcp Remove it, or run one shared HTTP server Fix works, measured
2 Helpers outlive the session that started them agent-slowdown orphans Stop those whose parent and session both died Fix works, read the list before --kill
3 The agent runs find / or rg x ~ agent-slowdown guard A hook that refuses it before it starts Fix works, live in Claude Code
4 Electron and desktop apps grow agent-slowdown apps Restart the app. There is no other cure Report only
5 The Linux memory limit is silently off agent-slowdown memlimit Start Claude in a user scope Fix works, measured on Linux
6 Spotlight and antivirus scan agent folders agent-slowdown exclusions .noindex folders; Defender exclusions are your call Spotlight measured, Defender not

Details and numbers: MCP copies · orphans · broad searches · desktop apps · Linux memory limit · indexers

The one that costs the most

one stdio MCP server, 26 sessions on a 32 GB Mac:      84 processes, 2.0 GB
one stdio MCP server, 56 sessions on a Windows box:   176 processes, 7.0 GB
the same kind of server over HTTP, any session count:   1 process,  42 MB

Summed memory over-counts pages shared between copies. The copy count is exact.

There is no lazy flag for this in Claude Code. Deferred tools do not save the process, because it has already started by then. Lazy means the server is not in the always-on config. The other half of the fix, running one shared server that survives reboots, has its own repository: mcp-daemon-diet.

Stop broad searches before they start

The hook needs the command to stay on your PATH, so install it first:

pipx install git+https://github.com/tonydzi/agent-slowdown
agent-slowdown guard --print-config

Paste the result into ~/.claude/settings.json. Blocked commands exit 2, and the agent reads the reason and retries with a narrower path. It reads the command as text and never runs it. It is a seatbelt against the accidental command, not a sandbox.

What this does not do

  • It does not make your agents faster. It removes what is slowing the machine down around them.
  • It cannot shrink a desktop app's memory.
  • It has no idea how many agents your machine can take. Nobody has a good number; our rough figure is 1 to 1.5 GB per active session with all its helpers.
  • Windows process reading runs in CI, but we have used orphans --kill on macOS and Linux test processes only.

Exit codes: 0 nothing found, 1 found something, 2 could not read the process list, 3 a config file could not be read. A check that could not look says so; it never reports clean.

Who made this

Built while running a fleet of coding agents across several machines, and paying for each of these lessons with a frozen computer. Claude wrote most of it. Codex and GLM reviewed it and found real bugs, which are fixed. Measurements are ours unless linked.

If it found something on your machine, an issue with the output helps more than a star.

TonyDzi, Palo Alto AI Research Lab · second brain, agent coordination, persistent memory: github.com/tonydzi · WhatsApp +1 341 222 9178

MIT license.

READ THIS WITH AI

One click and an agent reads the repo, pulls out the patterns and helps you apply them to your own work.

Codex - open ChatGPT - open Claude - open

Copy the prompt (works in any agent: Gemini, Grok, a local model, your own CLI)
Read this repo: https://github.com/tonydzi/agent-slowdown (“agent-slowdown” - Why parallel Claude Code, Codex, Cursor and Antigravity sessions make your machine crawl: per-session MCP copies, orphaned helpers, broad searches, and what actually helps). Work out what problem it actually solves, pull out the reusable patterns and help me apply them to my own setup. Start by asking what I am working on.

— TonyDzi, Palo Alto AI Research Lab · second brain, agent coordination, persistent memory: github.com/tonydzi

Metadata

Release files for agent-slowdown 0.1.1

For a detailed explanation of source distributions (sdists) and built distributions (wheels), please see the package formats documentation.

Source distribution (sdist)

Source distribution for agent-slowdown 0.1.1
File Size Uploaded Provenance Actions
agent_slowdown-0.1.1.tar.gz 40.0 kB With provenance Details for agent_slowdown-0.1.1.tar.gz

Built distribution (wheel)

Table of built distributions (wheels) for agent-slowdown 0.1.1
File Interpreter ABI Platform Actions
agent_slowdown-0.1.1-py3-none-any.whl Python 3 none any Details for agent_slowdown-0.1.1-py3-none-any.whl

Total release size: 69.6 kB

Release files / agent_slowdown-0.1.1.tar.gz

Download URL agent_slowdown-0.1.1.tar.gz
Size 40.0 kB
Tags Source
SHA-256 checksum
How to use checksums
3127b652ab70179e773e63a31137a51b7e635f775de900b764667d94a0d5e5aa
BLAKE2b-256 checksum
How to use checksums
dc75102b87108fb6bf7e694346031cfb8588e12337f0e6e3b46b7e7d28caa239
Upload date
Uploaded using Trusted Publishing?
What is trusted publishing?
Yes
Uploaded via twine/7.0.0 CPython/3.13.14

Provenance

Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.

PyPI Publish Attestation

PyPI verified that this artifact, at this checksum, originated from the publisher listed below.

Signed by GitHub Actions, verified by PyPI on Nov 10, 2026.

Transparency log

Release files / agent_slowdown-0.1.1-py3-none-any.whl

Download URL agent_slowdown-0.1.1-py3-none-any.whl
Size 29.6 kB
Tags Python 3
SHA-256 checksum
How to use checksums
a82d0133cfaba9b2832beaa94e74105a3c39d45e7870d9d475daee2eea8f669f
BLAKE2b-256 checksum
How to use checksums
13763380a5d67620a712f89757d0778719116bf4814ce0199c18100be60c2c09
Upload date
Uploaded using Trusted Publishing?
What is trusted publishing?
Yes
Uploaded via twine/7.0.0 CPython/3.13.14

Provenance

Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.

PyPI Publish Attestation

PyPI verified that this artifact, at this checksum, originated from the publisher listed below.

Signed by GitHub Actions, verified by PyPI on Nov 10, 2026.

Transparency log

Security for agent-slowdown 0.1.1

Provenance

Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.

Full provenance

Both files in this release have been uploaded with provenance metadata.

Select from 2 files

Attestation for agent_slowdown-0.1.1.tar.gz

PyPI Publish Attestation

PyPI verified that this artifact, at this checksum, originated from the publisher listed below.

Signed by GitHub Actions, verified by PyPI on Nov 10, 2026.

Transparency log

A note on package security

PyPI is a repository for distributing packages, not a service for auditing them. While third-party security researchers and our staff do their best to detect and remove malicious packages, PyPI cannot guarantee the quality, security or suitability of any given package.

To learn more about protecting yourself from malicious actors, visit our package security documentation.

Release history Release notifications | RSS feed

This release

0.1.1 This release

2 release files Full provenance

Anthropic, PBC Visionary sponsor Bloomberg Visionary sponsor Hudson River Trading Visionary sponsor Meta Visionary sponsor NVIDIA Visionary sponsor Microsoft Sustainability sponsor Depot Continuous Integration AWS Cloud computing and Security Sponsor Datadog Monitoring Fastly CDN Google Download Analytics Sentry Error logging StatusPage Status page