Skip to content
plugmcp
DirectorySkillsCategoriesBlog
EN/δΈ­ζ–‡
Sign inSign up
Home/Servers/Glitch Toolkit

Glitch Toolkit

I
by io.github.AbstractGlitch

Read-only checks that fail a guardrail which is present but has stopped refusing anything

Power your agent

Add this server

The publisher's own command. Not a PlugMCP endpoint.

Local command
uvx glitch-toolkit==0.1.4
Provider
Claude Code
claude mcp add io.github.abstractglitch-glitch-toolkit -- uvx glitch-toolkit==0.1.4

Run in a terminal.

Claude Desktop
{
  "mcpServers": {
    "io.github.abstractglitch-glitch-toolkit": {
      "command": "uvx",
      "args": [
        "glitch-toolkit==0.1.4"
      ]
    }
  }
}

claude_desktop_config.json

Step-by-step: add this in Claude

Remote URL or a command on your machine? How the two installs differ

Run on PlugMCP

Host a copy with our keys and logs.

A new server you create. It is not this listing's official endpoint. Use the publisher install when you want their URL and their credentials.

Launch

Related MCP

View more
  • FirecrawlπŸ”₯ Official Firecrawl MCP Server - Adds powerful web scraping and search to Cursor, Claude and any other LLM clients.
  • N
    NotionToken-auth Notion MCP: pages, databases, blocks, comments, files. Runs headless in CI and agents.
  • Context7Context7 Platform -- Up-to-date code documentation for LLMs and AI code editors

Related skills

View all
  • reaReverse engineer anything with agents, from app behavior down to native binaries.
  • rea-changelog-updatePrepare or rewrite REA release changelogs and GitHub release notes from pinned Git history, with verified contributor thanks and Release Please synchronization.
  • rea-tool-designDesign or change REA investigation tools, CLI/MCP contracts, provider capabilities, and Evidence semantics.

Read-only checks that fail a guardrail which is present but has stopped refusing anything

glitch

A check that has stopped refusing things still passes. That is the failure this looks for.

On 8 September 2026, in the repository this package was extracted from, a test guarded the one chapter of a paid book that is given away free β€” asserting the sample stops at its cut and does not leak the rest. It was green. It was searching a page that had no book on it, for passages it therefore could never find, and passing. Nothing was broken; the redaction worked fine. The alarm had been disconnected and was still showing a green light.

Every guardrail decays that way eventually, and the decay is silent, because a guardrail that has stopped refusing looks exactly like one with nothing to refuse.

glitch installs six guardrails into your repository and then, whenever you ask, runs each one against a case it is supposed to refuse. One that no longer refuses anything fails here, loudly, instead of passing quietly.

pip install glitch-toolkit

glitch install     # put the artifacts in this repository
glitch status      # what is installed, what is not
glitch check --all # verify every step

No dependencies, no network calls, no telemetry. It writes nothing outside the directory you point it at.

What it actually checks

A step is not complete because a file is present. Every check does three things and reports which of them failed:

  1. finds the artifact
  2. runs it, and expects it to work
  3. runs it against a deliberately broken case, and expects it to refuse

Without the third, a check passes the moment you copy a file in, whether or not that file has any teeth left, and it would be green for every reader forever. tests/test_cli.py exists to prove the checker fails that case; its SABOTAGE test installs artifacts that run, exit 0 and refuse nothing.

What install will not do

It will not write your CLAUDE.md, FLEET.md, FLOOR.md or PLAN.md. Four of the six steps are checked against your own file, because for those four the file is the work: a rules file that holds, a desk table with one committer, a floor measured twice, a plan someone else approved. A command that wrote them would turn the path into "you ran an installer". The FLOOR.md and PLAN.md templates ship deliberately unpassable for the same reason.

It will not overwrite. An artifact already in your repo is left alone and reported as kept; --force is how you say otherwise.

It writes nothing outside the directory you point it at, makes no network calls, and has no dependencies outside the standard library.

Running it without installing it

cli.py is one file and stays one file. Copy it into a repository and status and check work with nothing on the path and no install step β€” that property is deliberate and there is a test for the search paths it uses. Only install needs the rest of the package, and it says so plainly rather than failing oddly.

The MCP server (read-only)

uvx --from 'glitch-toolkit[mcp]' glitch-mcp --repo .

Install the extra in an isolated environment, not a system Python. The checker has no dependencies and that is a promise. The [mcp] extra is the opposite: mcp>=2.0 pulls pydantic, starlette, cryptography, opentelemetry and a dozen more, and pip will happily UPGRADE what is already there to satisfy them. Done against a global interpreter on 8 September 2026 it replaced pydantic 1.10 with 2.13 and starlette 0.46 with 1.6, breaking an unrelated FastAPI application on the same machine. uvx builds a throwaway environment and touches nothing else, which is why the registry entry launches it that way.

A virtualenv is equally fine:

python -m venv .venv && .venv/bin/pip install 'glitch-toolkit[mcp]'

It offers the checks to an agent as three tools β€” glitch_status, glitch_check, glitch_ledger_tail β€” and appends every question and answer to .claude/toolkit/ledger/ledger.jsonl.

It gates nothing. It cannot pause, block, refuse or intercept any action. It has no database connection, no credential and no network call. That is the whole first version, on purpose: a server that stands between an agent and a production database is serious software, and the honest order is to run read-only first, read the ledger, and find out what it would have refused before giving it the power to refuse. A gate built before that record exists is a guess with permissions.

The checks write nothing into your repository. The server breaks that in exactly one place β€” it appends to the ledger β€” and --no-ledger turns off even that, at the cost of the only thing worth keeping.

In Claude Code, .mcp.json:

{
  "mcpServers": {
    "glitch": { "command": "glitch-mcp", "args": ["--repo", "."] }
  }
}

The ledger

Append-only JSONL. Nothing rewrites a line it did not just write; a record that is overtaken is superseded by a new one and both stay; a field nobody measured is null rather than 0; a half-written last line is skipped and counted, never repaired, because repairing it means rewriting the file.

Tests

python tests/run_all.py      # all three suites, 40 tests

The server suite skips cleanly without the [mcp] extra and the runner reports that as SKIPPED rather than passing, because a green line meaning "we did not look" is the exact failure the gate-check practice exists to catch.

Where the checks come from

Every check here exists because something went wrong, and CORPUS.md is the list β€” what happened, what it cost, and which check catches it now.

It also names the ones nothing here catches yet, including the two that cost this package a version number each. A corpus that recorded only its solved failures would be doing the thing this package is about.

Licence

Apache License 2.0 β€” see LICENSE and NOTICE. Chosen over MIT for the patent grant.

Everything in this repository is under it. Use it commercially, change it, redistribute it.

The links above are absolute on purpose: this README is also the package's description on PyPI, where a relative link resolves against pypi.org and returns a 404.

Where this comes from

This repository is a published mirror. The package is developed inside a private monorepo alongside the book Building Your Store Or Your SaaS With Claude, whose practices it installs and checks, and it is pushed here as a subtree. The book, the shop that sells it and the rest of that repository are not open source and are not here. Nothing is being withheld from this repository that belongs to the package.

Issues and pull requests belong here rather than there, because here is the part anyone can read.

Status

Version 0.1.1. It installs, and the practices it checks are the six the book argues for. 0.1.0 was the first release; 0.1.1 changes this description and adds continuous integration, and nothing about what the code does.

What it is not yet: it gates nothing. glitch-mcp reports and records and cannot block an agent from doing anything. That is deliberate and the reasoning is in mcp_server.py β€” a server that stands between an agent and a production database should earn its evidence before it earns the power to refuse.

How do I install Glitch Toolkit?
Copy the stdio config for Cursor, Claude Desktop, Claude Code, or VS Code on this page. It uses the publisher's install, not a PlugMCP proxy.
Does Glitch Toolkit need an API key?
No required environment variable is listed on this record.
What transport does Glitch Toolkit use?
Glitch Toolkit runs locally over stdio.
When was Glitch Toolkit last checked?
Last checked 2026-10-16. Read-only checks that fail a guardrail which is present but has stopped refusing anything

No tool list yet. Remote servers get one after the first health check.

plugmcp

Plug any tool into any agent. MCP servers for people who do not write servers. Gateway for volume, n8n for apps.

  • github.com/TODO
  • x.com/TODO
  • linkedin.com/company/TODO

PlugMCP, 2026

Product
  • Browse
  • Install
  • Host your own
  • Voice mode
  • Browser
Directory
  • All MCP servers
  • Agent skills
  • Categories
  • New
  • Voice
  • Submit a server
Developers
  • Docs
  • Blog
  • What is MCP
  • What is WebMCP
  • Status
Company
  • About
  • Contact
  • Terms
  • Privacy
  • DPA