Check
GOL Productions builds deterministic gates for the new era of modern technology: checks where the answer comes from your own machine, never from a guess. Check connects AI to reality before it acts. It is our gate for AI coding agents.
1. Get your key at the console: sign in, press Enable authenticator and scan the QR code with an authenticator app, then press Create GOL API Key, and top up your balance there.
2. No Node.js? Install Node.js 18 or newer from nodejs.org (free).
3. Run one line in a terminal, yourself, not through your AI, and paste your key when it asks:
npx @golproductions/check@latest --install
It asks for your key (masked), lists the AI agents it found, and adds Check to them when you press Enter.
4. Restart your AI agent.
Windows or Linux; macOS isn't supported yet. The installer never makes keys and never prints yours. Run by an AI with no key on the machine, it installs nothing and tells you to run it yourself.
The install guide covers making it work in each agent, what to expect, and how to know it's working.
What does Check do?
Before your AI agent runs a command or writes a file, Check asks your own machine whether it will work. Does the program exist? Is the package real? Will the file still parse? If your machine proves it won't, Check blocks it and shows the AI your machine's answer, so its next attempt is a fix instead of another guess.
At the end of every reply you get a one-line receipt:
Check: 7 actions checked, 1 blocked (pkill: not found)
Why?
AI agents guess. They call programs that aren't installed, install packages that don't exist, use git branches that were never made, and write code that won't parse. Each guess costs a failed run, a turn spent reading the error, and another guess, and some mistakes stay hidden until the file is run later.
Your machine already knows the answers. Check asks it first.
Examples
Three real blocks, on a Windows machine running Git Bash.
A Linux tool, with its error hidden. Stopping a dev server, Claude reached for pkill, which Git Bash doesn't have, and sent its errors to /dev/null, so the failure would never have shown.
taskkill //F //IM workerd.exe >/dev/null 2>&1; pkill -f "wrangler dev --remote --port 8799" 2>/dev/null; ...
check: bash on this machine said: type: pkill: not found
An empty answer that would have read as "none". In a site check, Claude asked dig for DNSSEC records and labelled empty output as "none". With no dig installed, the output is always empty, so the site would have been reported as having no DNSSEC, whatever the truth.
dig +short DS golproductions.com @1.1.1.1; echo "DS ^ (empty = none)"
check: bash on this machine said: type: dig: not found
Code that Git Bash would break. On Windows, Git Bash halves the backslashes in code passed to node -e before node sees it. The regex Claude wrote was valid; the one that would have arrived was not.
node -e 'const s=require("fs").readFileSync(process.argv[1],"utf8");const m=s.match(/\\n\s*#[^\\]*/g)...'
check: Git Bash will break the code passed to node -e before it runs: it turns \\ into \ on its command line,
and the code that arrives does not parse (node said: Invalid regular expression: missing /).
Put the code in a file with the Write tool and run the file.
Each time, the AI read the answer and fixed it on the next try: it dropped pkill and kept taskkill, dropped the dig line, and moved the code into a script file.
What it checks
| Question | Who answers |
|---|---|
| Does this program exist? | your shell (type) |
Does this ./script exist? | your file system |
| Is this git subcommand, branch, tag, commit or tracked file real? | your git |
| Does this package exist (npm, PyPI, crates.io, RubyGems, Packagist, NuGet, Homebrew, Chocolatey, PowerShell Gallery)? Does the version asked for? | the package registry |
Does the module python -m would run exist? | your Python, from the folder the command runs in |
Is the package a download-and-run tool would fetch real (npx, pnpm dlx, yarn dlx, bunx, uvx, pipx run)? | your machine (is it installed here?), then the npm registry or PyPI |
Is the Docker image or tag that docker pull, run or create would fetch real? | your Docker (is it here?), then Docker Hub |
Does the folder cd goes to, or the npm, pnpm or yarn run script exist? Does the git push remote? | your file system, package.json, your git |
Does the script node, bash, sh, ruby or php would run exist? Does the file cat, head, tail, wc, ls, Get-Content or Get-ChildItem would read? | that tool itself (node --check, bash -n, ruby -c, php -l, or the reading tool with its output thrown away), in its own words |
| Does the command parse? | your shell's own parser |
Does the code passed to node -e or python -c, or fed to node or python - through a heredoc, parse as it will arrive? | node and python, including after Git Bash halves its backslashes |
Will this file still parse after the write, whether written by Write, Edit, a heredoc, echo or printf? | bash -n, Python, PowerShell, node --check, and the JSON parser npm uses for package.json |
| Does a package a JavaScript or TypeScript write starts importing exist? | your project (installed or declared?), then the npm registry |
| Does a TypeScript write use a function, method, export or option a library actually has? | your project's own TypeScript (5 or 6), against the library's own type definitions; new errors in the edit only, capped at 2.5 s |
Commands are read by your own bash (PowerShell commands by PowerShell's parser) without running any of them. Bash also reads both sides of || and else, the commands inside $(...) and backticks, and the code given to bash -c or sh -c. JavaScript writes: .mjs and .cjs files are always checked. A .js file that parsed before the edit is blocked only if the result is valid as neither a CommonJS script nor an ES module, so bundler code passes. A new .js file is blocked only for an error JSX, Flow, TypeScript or decorators never cause (an unclosed bracket, an unexpected end of input), so those dialects pass. Other .json files, such as tsconfig.json with comments, are not checked.
When a Bash command fails because Git Bash changed its inline code, Check adds a note telling the AI what arrived and what to do instead.
Where Check stays silent
Where your machine can't give a definite answer, Check lets the action through rather than guess:
- commands run through
eval,env,xargs,sudo, or a login or interactive shell (bash -l,bash -i) - git and
./scriptchecks after a command changes folder (cd,pushd,git -C) - everything after
PATH=,sourceor.in the same command - programs after an install in the same command (
npm i -g x && x) - a script run after an earlier part of the same command, which may have made it
- a program or file the command tests for first (
command -v,which,type,hash,test,[; in PowerShellGet-Command,Test-Path,try/catch,$?,$LASTEXITCODE) - missing programs in PowerShell commands that load a module, define an alias or function, dot-source a script, change
$env:Pathor run on another machine, or that both installed PowerShells don't agree on - a
python -mmodule after the command changes folder; arubyorphpscript with options before it; scripts run by Python, Perl or tsx (none of them can read their own arguments without running code) - files written by
echoorprintfin any form but one single-quoted argument with>;.jsxand TypeScript files; text a parser can't tell from a dialect (an unterminated string in a new.jsfile) - package checks for an ecosystem this machine installs from a private registry, or a command that names another source (
choco -s,Install-Module -Repository,uvx --from) - version ranges and wildcards (
~> 7.0,^2.0,13.*): the name is still checked; versions on the PowerShell Gallery; winget, scoop and Go modules - Docker images from another registry (
ghcr.io/...), by digest, or under a Docker Hub account that exists (a private repository looks missing from outside) - imports in projects with path aliases, workspaces or import maps, and Python imports (an import name is often not the package's name)
- library names in TypeScript 7 projects (no programming interface), in libraries without type definitions, in plain JavaScript, in Python, and whenever the project's TypeScript takes longer than 2.5 s (the write goes ahead)
It also cannot know intent: a real command that does the wrong thing passes. Check is not a security tool.
Fails open
Server unreachable, rate limited, balance empty, key rejected: the command runs, with a warning at most once a day. Check never blocks because Check itself failed, and your balance never blocks your agent.
How it hooks in
| Claude Code event | What Check does |
|---|---|
| PreToolUse (Bash, PowerShell, Write, Edit, NotebookEdit) | the checks above |
| PostToolUse (Bash, PowerShell) | after a Cloudflare Pages deploy or an npm publish: proves it landed (below); anything else returns at once |
| PreCompact, SessionStart (after compaction) | the compaction note (below) |
| PostToolUseFailure (Bash) | the Git Bash note, only when it applies; free |
| UserPromptSubmit | counts the message for billing (its ID only); prints nothing |
| Stop | the one-line receipt |
Other agents
The installer adds Check to each agent's own user-level hook file, keeping your hooks, and --uninstall removes only Check's entries.
| Agent | Hook file | Commands | File writes | Message billing |
|---|---|---|---|---|
| Codex CLI | ~/.codex/hooks.json | PreToolUse (Bash) | apply_patch that adds a file | UserPromptSubmit (turn ID) |
| Cursor | ~/.cursor/hooks.json | beforeShellExecution | not checked (no hook before an edit) | beforeSubmitPrompt |
| Gemini CLI | ~/.gemini/settings.json | BeforeTool run_shell_command (Linux) | write_file, replace | BeforeAgent |
| Windsurf / Devin Desktop | ~/.codeium/windsurf/hooks.json | pre_run_command | pre_write_code | pre_user_prompt |
| GitHub Copilot CLI | ~/.copilot/hooks/check.json | preToolUse (bash, powershell) | not checked (tool arguments undocumented) | userPromptSubmitted |
On Windows, Codex, Cursor and Windsurf / Devin run commands in PowerShell, so they're checked as PowerShell; on Linux every agent's commands run in bash. Gemini CLI's Windows shell is undocumented, so its commands are checked on Linux only. Codex asks you to trust a new hook once: run /hooks in Codex and approve Check. The receipt at the end of a reply is Claude Code's only. Not supported: Cline (its hooks are now code plugins, not scripts) and any agent with no hook before an action.
The connector: any other environment
The connector is how Check is integrated, the same for every tool. If your environment has a hook that can run a program before the AI acts and stop the action, install the gate on its own (npx @golproductions/check@latest --install --manual: key and gate, no agent changed), point the hook at it (node ~/.check/check-hook.mjs) and send it one JSON object on stdin:
{ "check": 1, "event": "run", "session": "chat-42", "shell": "bash", "command": "npm i [email protected]", "cwd": "/path/to/project" }
event is message (with prompt), run (with command and shell: bash or powershell), write (with file_path and content) or edit (with file_path, old_string, new_string). Each message is billed once; every check after it in the same session belongs to it. The message text never leaves the machine.
The gate answers {"decision":"allow"}, or {"decision":"deny","reason":"check: ..."}: stop the action and give the reason to the AI. An allow may carry a notice for the person. Anything other than deny, including no answer, means go ahead. Full contract: https://golproductions.com/guide#connector
Check never edits your code. The one file it adds is the stamp in a Cloudflare Pages deploy folder (below).
Proves deploys and publishes
Your AI says "it's live" or "it's published" without checking. Check checks.
- Before
wrangler pages deploy <folder>: Check stamps the folder: it writes/.well-known/jogga.jsoninto it, listing every file's path and SHA-256 and a deploy ID. A deploy Check can't stamp is stopped, with the reason. Run the deploy as its own command (acdbefore it is fine), so nothing can change the folder after it's stamped. - After the deploy: Check reads the address Cloudflare printed and our server looks for that stamp there, for up to 15 seconds. This deploy's stamp, with every file matching: live. Another deploy's stamp, or files that differ: not live. No stamp at all: unconfirmed, which means Check cannot say it is live.
- After
npm publish: Check packs the folder again (no scripts run) and checks it against the file npm serves, byte for byte.
The AI, and you, see the answer straight away: Check: LIVE: ... All 112 files match, byte for byte., Check: NOT LIVE: ... or Check: UNCONFIRMED: ... doesn't serve this deploy's stamp.
It landed: free. It didn't (not live, or not published as built): charged like a block. Unconfirmed, or couldn't tell (npm still listing, a server down): free. Proves Cloudflare Pages deploys and npm publishes to the public npm registry, tested on Windows. Other deploy commands aren't checked yet.
Remembers through compaction
When Claude Code compacts a long conversation, your AI keeps only a summary, and a summary paraphrases. Right before compaction, Check asks our server how to write a note, and your own machine writes it from its own record of the conversation: every message you sent, word for word, the files changed, the last commands run, and the AI's last reply. The note is saved on your machine, in ~/.check/handoff, for 30 days. Right after compaction, the AI is handed your newest messages from it, word for word, and told where the full note is.
Nothing from your conversation is sent: our server sends instructions only, and the request carries your key and your network address, like any request, and nothing from the conversation. (Versions of the gate before 5.3.0 sent your messages to our server to write the note; the gate updates itself within an hour.) A$0.12 a note, from your balance, charged when Check asks for the instructions. Without enough credit, there's no note, nothing is charged, and the AI is told why.
What leaves your machine
Per message: its ID, never its text. Per command: the command text with credential patterns redacted (API keys, tokens, JWTs, private keys, Bearer tokens, emails, card-like numbers; a secret in a PowerShell $env: assignment is sent as written), the working folder with your username masked, and your machine's answers. To know which AI tool is asking: the layout of its hook message (field names and lengths, never their values). Per file write: the file's name, its size and whether it already exists; if a parser rejects it, the parser's one-line error message (redacted the same way), which can quote the word it stopped at; for a JavaScript or TypeScript file, the names of npm packages the write starts importing, with whether each is installed or declared in the project; and, when your project's own TypeScript says the write uses a name a library doesn't have, the library, that name, the error code, its line and TypeScript's one-line message. File contents never leave. --feedback sends your message, Check's version, your OS, which agents Check is installed in and a shortened form of your key. What is kept: a record of each check (time, a one-way hash of your key, your network address, Check's version, the AI tool and the verdict) and your billing record. With improvement records on (the default for keys created from 14 October 2026, switchable at the console), also the command with credentials removed, the name of a file being written, and the reason Check gave. Never your file contents, and never your messages. After a deploy: the deploy's address and ID; our server reads the stamp (file paths and SHA-256 fingerprints, never contents) from your deployed site. The stamp is a file you publish. After an npm publish: the package's name, version and fingerprint. Details: privacy policy.
Pricing
A block is one action Check stops before it runs, or a deploy or publish Check proves didn't land. A deploy it can't confirm, for lack of a stamp, is free. Each block is A$0.12 from a prepaid balance at the console. Everything Check allows, and every deploy or publish that landed, is free. A compaction note is A$0.12. When the balance runs out, Check stops checking and your commands still run. No subscription, nothing auto-recharges, credits don't expire. Full terms: golproductions.com/terms.
Keys
One kind of key: the GOL API Key. It's made only at the console, by a signed-in account with two-factor authentication, and it draws on its prepaid balance. On a machine it's saved in ~/.check/key: the install asks for it, or set it yourself with --key.
Keep it safe: never paste it into an AI chat or prompt, a project file, a commit or a screenshot. Anyone with it can spend your credit. If it leaks, press Rotate at the console: the old key stops at once and the balance moves to the new one. On a server or in CI, pass it through the GOL_API_KEY environment variable from your secret store.
Getting your key back out
Check never prints the full key: every surface shows it masked, because it is a credential and an AI agent reads everything a hook prints back.
npx @golproductions/check@latest --print
Run this yourself, in a real terminal. The first time, it asks you to set a password (never written to disk, never told to any agent) and saves your key encrypted to a file on your Desktop. Every time after, it asks for that password before showing the key. Paste it into "Connect key" at the console while signed in to link it to your account, then delete the encrypted file.
Manage
npx @golproductions/check@latest --key # set your GOL API Key (run it yourself)
npx @golproductions/check@latest --status # key + installation status
npx @golproductions/check@latest --credits # balance and price
npx @golproductions/check@latest --print # get your key back, password-gated
npx @golproductions/check@latest --uninstall # remove hooks
The gate keeps itself current, checking for updates in the background at most once an hour; re-run the install only to update the installer itself.
Terms
By installing or using Check you agree to the Terms of Service and Privacy Policy. The service is provided as is; see the Terms for the full picture, including your Australian Consumer Law rights.
Product · How it works · Pricing · Console
Trademarks
Not affiliated with, endorsed by, or sponsored by Anthropic. Claude and Claude Code are trademarks of Anthropic, PBC. Cursor, Gemini, Codex, Windsurf, Devin and GitHub Copilot are trademarks of their respective owners. They are named only to say which tools Check works with. GOLProductions is an independent company.
License
Copyright (c) 2026 GOL Productions. All rights reserved.