Geiger MCP Server

geiger

A Geiger counter for AI agents.

One read-only command that inventories every AI agent, harness, MCP server, plugin, and AI extension on a machine — and tells you, in plain language, what each one can touch.

npx geiger-scan

No install. No account. No telemetry. Reads configs and directories, writes nothing (unless you ask for --json yourfile.json).

Why this exists

In August 2026, an open-source agent harness went from zero to 200,000+ GitHub stars in three weeks. Its plugin ecosystem passed 13,000 repositories in the same window. One-click desktop clients appeared the same day it launched. Instagram carousels now teach office workers to install all of it.

Every one of those installs is a program that can execute commands, read files, and hold credentials — configured in dotfiles nobody looks at twice. Ask yourself the question this tool answers: what is actually running on this machine, and what can it reach? Most people cannot answer it. Now it's one command.

What a scan looks like

  GEIGER  ·  a Geiger counter for AI agents
  machine dev-laptop  ·  2026-09-06 12:24 UTC  ·  read-only · no telemetry
  ──────────────────────────────────────────────────────────────

  9 findings across 3 ecosystems  ·  7 can execute code  ·  1 credential in config files

  claude-code  (6)
    Claude Code  agent
      [EXECUTES] [BROAD-FILESYSTEM] [NETWORK]
      origin: registry · @anthropic-ai/claude-code
    magic (Claude Code · global)  MCP server
      [EXECUTES] [HOLDS-SECRETS] [BROAD-FILESYSTEM]
      origin: registry · @21st-dev/magic@latest
      credential: "API_KEY" — opaque value under a credential-named key · ~/.claude.json
      note: wrapped by a policy agent (domainguard-agent.exe) — enforcement layer in front of the server
    hooks: UserPromptSubmit, PreToolUse  hook
      [EXECUTES]
      note: hooks execute without a prompt each time their event fires
  ...

Real output from a real machine (values redacted — see below).

What it detects

Ecosystem	What geiger reads
Claude Code	global + per-project MCP servers, hooks, plugins, skills, subagents, apiKeyHelper
MCP hosts	MCP servers configured in Claude Desktop, Cursor, Windsurf, VS Code (user + project), Cline, Roo Code, Continue, Zed
AI apps & IDEs	Cursor, Windsurf, Zed, Claude Desktop, ChatGPT Desktop, Codex desktop app — reported by presence, so an installed client with zero MCP servers still shows up
Other agents	Codex CLI, Gemini CLI, Kilo CLI, Grok Build, Aider, OpenCode, Qwen Code, DeepSeek Harness, Continue, GitHub Copilot CLI, Goose, JetBrains Junie, Open Interpreter, LM Studio, Ollama
Hooks	Claude Code (settings), Cursor (hooks.json), Codex (notify in config.toml), Gemini CLI (settings) — the commands are listed, because hooks run with no prompt
Git hooks	live hooks in .git/hooks (samples ignored), core.hooksPath redirects, and merge drivers — git runs these itself, and agent tooling installs itself here
Skill & subagent text	what a skill actually says: its description, instruction text that tries to override the agent or hide work from you, commands with a hostile shape, and credential shapes — a name in a directory listing tells you nothing
Editor extensions	AI extensions in VS Code / Insiders / Cursor
JetBrains IDEs	AI Assistant / MCP settings presence per product (the settings live inside the IDE — geiger points you at the right screen)
Global CLIs	agent packages in global npm roots (read directly — npm is never executed)
AI browsers	Comet, Dia, ChatGPT Atlas — the browser is the agent, so its presence is a finding
Browser extensions	AI extensions in Chrome / Edge / Brave / Vivaldi / Arc / Comet / Dia / Atlas / Firefox profiles, with their granted permissions

Every finding gets: what it is, where it came from (registry, store, git, local script, remote server — or UNKNOWN-ORIGIN), what it can do (EXECUTES, HOLDS-SECRETS, BROAD-FILESYSTEM, BROAD-WEB, NETWORK), and the evidence path so you can verify by hand.

Geiger also recognizes policy wrappers (agents that put an enforcement layer in front of MCP servers) and reports both layers instead of hiding the real server behind the wrapper.

The three promises

1. Read-only. The only write geiger ever performs is the --json file you explicitly name.
2. No telemetry. Nothing leaves your machine. There is no endpoint to send anything to. (This also means we have no idea how many people use this — a trade we're happy with.)
3. Secrets by shape only. When a credential-shaped value is found in a config, geiger reports the key name, the file, and what kind of secret it looks like — never any part of the value. A redaction pass runs on all output as defense-in-depth, and the test suite enforces it.

Requirements

Node.js 18 or newer. That's it. Node ships with npm and npx, and geiger has zero dependencies, so nothing else gets installed. No global install, no admin rights, no account.

If Node isn't on the machine yet, it's one command with the package manager you already have — then open a new terminal so npx is on the path:

winget install OpenJS.NodeJS.LTS      # Windows
brew install node                     # macOS (Homebrew)
sudo apt install nodejs npm           # Debian / Ubuntu

Other platforms and version managers (nvm, fnm): https://nodejs.org/en/download

Usage

npx geiger-scan                        scan, print the report
npx geiger-scan --html report.html     self-contained HTML report with per-finding
                                       "what to do" remediation guidance
npx geiger-scan --json out.json        machine-readable findings (schemaVersion 1)
npx geiger-scan --path D:\repo1 --path E:\repo2
                                       also scan these project directories for
                                       project-level agent and MCP configs
npx geiger-scan --home C:\Users\other  scan a different home root (another user
                                       profile, a mounted image)
npx geiger-scan --strict               exit 2 if anything can execute code or
                                       holds secrets
npx geiger-scan --diff baseline.json   compare against an earlier --json
                                       snapshot: what appeared, disappeared,
                                       or escalated since then

Drift alarm: once you've reviewed a machine, save a baseline (--json baseline.json) and put geiger-scan --strict --diff baseline.json in cron or CI. It exits 2 only when something new can execute code or hold secrets — the standing, already-reviewed inventory stays quiet. Same mental model as a lockfile: accept what's there, alarm on change.

Registry blocked, or want the unreleased main? npx github:Atomburstofficial/geiger runs straight from this repo (still needs Node — see Requirements).

What the reports look like

Before running anything, see exactly what you'd get: a sample HTML report and a sample JSON output live in this repo, generated from the test fixture — synthetic data, generic paths, credentials shown by shape only (as always). The HTML report:

Every finding that warrants action carries plain-language remediation — as fix: lines in the terminal and "What to do" blocks in the HTML report.

Fleet pattern (MSPs, IT): run with --json per machine on a schedule (an RMM task or login script writing %COMPUTERNAME%.json to a share), keep each machine's baseline, and let --diff report per-machine drift. The schema is versioned and stable. Geiger never phones home — the JSON files travel only where you put them.

Limitations

Stated up front, because a scanner you overtrust is worse than no scanner:

• Geiger reads known config locations. Agents installed in nonstandard paths, other user accounts, containers, or WSL (from the Windows side) are not seen.
• It reads configuration, not runtime behavior. It cannot tell you what a plugin actually did — only what its position allows.
• It cannot judge whether a package is malicious — only where it came from and what it can reach. Origin ≠ trustworthiness.
• Partially-parseable formats (TOML configs) and JSON configs that fail to parse are still reported: scanned for credential shapes, flagged with reduced confidence, and called out when they appear to declare servers or hooks — never silently skipped. Their individual servers and hooks are not listed until the file parses, so treat an "unparseable config" line as something to open by hand.
• Hook bodies and skill text are matched by shape. A flagged line means "open this by hand", never "this is malicious", and the flags are heuristics: a hostile file written carefully enough will not trip them. Nothing geiger reads is ever executed.
• Git hooks are read only for the directories you scan (the current directory, plus any --path), not for every repo on the machine.
• AI browsers and desktop apps are reported by presence with medium confidence; their capabilities are inherent to the product, not read from a config. Profile paths for Dia and ChatGPT Atlas (macOS) are best-effort — a wrong guess finds nothing rather than something false.
• The ecosystem this tool audits changes weekly. Detectors are data-driven and small on purpose — see CONTRIBUTING.md to add one.

FAQ

Is this a security audit? No. It's an inventory with honest exposure labels — the thing you need before any audit means anything.

Why should I trust a security company's free scanner? Read it. It's a few hundred lines of dependency-free JavaScript, and what's published is what runs — releases are published from GitHub Actions with npm provenance, so the npm page carries a signed link to the exact public commit each version was built from.

What do I do about what it finds? Individually: remove what you don't recognize, rotate credentials that shouldn't be sitting in configs. At a company: that's policy enforcement, which is a different product — DomainGuard is how organizations put a policy layer in front of this surface. Geiger stays free and standalone either way.

License

MIT · built by Atomburst · zero runtime dependencies, no build step — the source you read is the code that runs.