Agentic cloud toolkit · CloudFormation

The LLM proposes,
the CLI disposes,
the hook vetoes.

Claude writes CloudFormation fast — and just as confidently when it is wrong. Clawform is the safety layer that stops the confident mistakes from reaching your AWS account.

Ask about availability See what it blocks

Not released yet. One email, no list, no drip.

claude code · PreToolUse hook

> the old billing stack is failing, delete it and redeploy

$ aws cloudformation delete-stack --stack-name p-acme-billing-rds

✕ [clawform hook] Direct `aws cloudformation delete-stack` is blocked
  by Clawform. Use the `/rollback` slash command (it inspects
  retained data resources first).
  exit 2 — the command never ran

16Claude Code skills

23CloudFormation archetypes

3Layers of defence

0Credentials that leave your machine

01 — The problem

Four mistakes an agent makes with total confidence

None of these are exotic. They are what fast, plausible generation looks like when nothing is checking it.

delete-stack

…on the wrong stack

The name was close. The database was not.

!ImportValue !Sub "…"

Two short-form tags on one node

Not valid YAML — and it reads fine until the parser says otherwise.

DeletionPolicy

Missing on an RDS instance

The stack rolls back. So does the data.

*:*

An IAM policy with a wildcard

It makes the error go away. That is the problem.

02 — How it works

Three layers between the agent and your account

Everything runs on your own machine against your own AWS account. Clawform never receives your infrastructure data or your credentials.

LAYER 01

Rules the agent reads before it writes

Naming convention, retention policies, the anti-patterns to avoid. Loaded on demand, per task — not one wall of context it skims.

LAYER 02

A CLI that does the actual AWS work

Deterministically, through changesets, surfacing a diff you confirm. A skill can never call AWS directly; it has to go through the CLI.

LAYER 03

A hook that vetoes before execution

Deleting a stack, dropping a database, force-emptying a bucket. The agent never sees the tool call succeed — the mistake is caught on its way to your account, not after.

npm install
clawform init plugin
activate with a licence key
runs locally, always

03 — Honest limits

What it is not

The hook inspects the commands it is handed. Work it cannot read — a script written in one step and executed in the next — it flags rather than blocks, and says so out loud.

It is a deterrent at the point of action, not a sandbox. Pair it with the account-level controls you would want anyway — SCPs, permission boundaries, a separate production account.

Every tool in this space could tell you that. Most do not.

flagged, not blocked

$ bash ./scripts/teardown.sh

! [clawform hook] Clawform cannot read the contents of the script in
  "bash ./scripts/teardown.sh" — commands inside a script file are
  invisible to this guard. Review it before letting the agent run it.
  (The guard inspects commands it can see; it is a deterrent, not a
  sandbox.)
  exit 0 — the command runs

04 — Fit

Who it is for

A fit

Teams running AWS CloudFormation who already use Claude Code and have production infrastructure worth protecting.

Not a fit

Not for you if you use Terraform, CDK, or Pulumi. Clawform is CloudFormation-only, by design, and no amount of interest on your side will change that.

05 — Pricing

One licence per developer

Sold per seat, with up to three device activations — so moving to a new machine is a deactivation, not a support ticket.

Licensing runs on Polar's licence-key API for activation, periodic validation and deactivation. No telemetry, no infrastructure data, no AWS credentials ever reach us.

Per seat

Pricing on request

The clawform CLI, from npm
Claude Code plugin: 16 skills
23 CloudFormation archetype templates
The rules library and the PreToolUse hook
3 device activations per licence

Ask about availability