DataSnack · the write gate for Claude and Codex

Rules for what Claude and Codex can send. Enforced by code.

DataSnack checks the writes your agent makes to outside services (e.g., Gmail, LinkedIn, Google Sheets, Stripe) against your rules, and code makes the call, not another LLM.

Same call, same rules, same answer. Every time.

Start free See how it works

ops-agent · CheckThree writes from this morning's run
Example
Google Sheets: add a row to "Leads"Cleared by the rule private-record · written
sent
LinkedIn post: "Three things we learned inspecting 40 buildings…"Held: a post to the public is asked about · waiting for you
held
Stripe refund of $480 to cus_8Hq…Refused by the rule no-refunds-over-200 · not sent
refused
Every write records what it was, which rule answered, and what the agent did next

How it works

Between your agent and everything it can write to.

Not email only: a message, a public post, a payment, a row in a shared sheet, a calendar invite, a delete. If it changes something outside the conversation, it is checked.

  1. Connect it once

    Add DataSnack as a connector in Claude, or install the hook in Claude Code and Codex. A few minutes, no code.

  2. Watch in Observe

    Nothing is stopped. You see every write, and what each rule would have done with it, while DataSnack drafts your rules.

  3. Switch on Check

    One switch on the dashboard, per connector. Now what your rules do not clear waits for you, and what they refuse never goes.

The architecture, one call from start to receipt

Why code

A line in a prompt is not a guardrail.

"Never email a customer without asking" works until the day the model reads it differently. The alternatives today are trusting the prompt or approving every call yourself.

How you gate your agentWho decidesSame call, same answer?What you do
A line in the promptThe modelUsuallyFind out afterwards
Approve every call by handYouYesClick yes all day, or stop reading
DataSnackCode, over rules you can readYesAnswer only what no rule covers

What "code decides" means

Facts are counted. Verdicts are computed.

Rules are expressions

The recipients, the links and where they go, the amount, how many went out today: code reads them from the call itself. Never-rules run first, then ask-me, then send.

A model may read, never decide

For what only a reader can tell (an invented number, a promise), Jev, a fast reading model, scores the text. Code compares the score with the rule's limit.

The same draft, the same verdict

What was read is kept, so asking twice gets the same answer. Anything too close to a limit, or that cannot be checked, is held for you rather than guessed.

No rules to write

It drafts the rules. You confirm them.

Sensible from the first call

Every tool starts with starter rules: payments, deletes, public posts and mass messages are asked about. Nothing to set up.

From instructions you already wrote

Paste a task's instructions and its "never" and "at most" lines become rules. They go live when you say they are right.

From what you approve

Your answers and marks become proposed rules, each shown with its evidence. One that contradicts any answer you gave is never shown.

No rule goes live without you: that is the point of a gate you can trust.

Observe, then gate

An observability tool on day one. A gate when you say so.

Each connector has its own mode, switched from the dashboard. Nothing changes in your agent.

Observe

Everything goes ahead. DataSnack records each write and what it would have held and why. Nothing is ever stopped, an outage included.

one switch

Check

What your rules clear goes out, what they do not waits for your answer on your phone, and what they refuse never goes.

With the hook, in Claude Code and Codex, a blocked call cannot run: the client stops it, whatever the model wants. In Claude's apps and Cowork, the agent asks first because your task tells it to, and DataSnack shows you when it did not. Handled, coming soon, has DataSnack do the sending itself.

One place

Your rules live in one place, not in each assistant.

Claude, ChatGPT, Codex and Claude Code each have their own settings, and none of them share. DataSnack sits outside all of them: connect an assistant and it gets the same starter rules, the same Outbox and the same record as the rest.

One set of instructions

Give a task the same name wherever it runs and its rules are drafted from the same instructions, for whichever tools that assistant uses. You confirm them, as always.

One place to say yes

Whatever was held, from whichever assistant, waits in one Outbox, and every write lands in one record.

Switch without starting over

Try a new assistant next month. Your rules were never in the old one.

A rule you add for one tool stays with that tool. The same service under another name in another assistant starts from the starter rules and your task's drafted rules.

What it checks

Any write, through any connector.

Every action is described the same way whatever app it goes through, so rules read the same everywhere: the starter rule "ask me about any payment" holds a Stripe refund and a PayPal payout alike.

See what your agent writes. Then decide what it may.

Connect it, start in Observe, and switch on Check when the rules look right. Free while in preview.

Start free Read the docs