Secure the Terraform Claude Code writes

The agent that wrote it is the fastest reviewer

Static scanners run after the code lands. When the author is an agent, you can move the review left of the pull request entirely: the same session that generated the Terraform scans it, reads the findings, and fixes them before a human ever sees a diff. That only works if the findings are precise — an agent drowning in false positives will "fix" benign code and churn. audytx's cross-resource context is what keeps the signal clean enough to hand an agent.

Two steps

1. Register the MCP server

   Get a free Client ID from the dashboard (the plaintext UUID is shown once), then add the server:

   claude mcp add --transport http audytx https://audytx.com/mcp \
     --header "X-Client-ID: YOUR_CLIENT_ID"

   That's the standard remote-HTTP MCP registration. audytx exposes scan_terraform, autofix_terraform, scan_cloudformation, and four more tools; discovery is open, and tools/call is authenticated by the Client ID you just passed.

2. Give Claude Code a standing instruction

   Claude Code reads CLAUDE.md at the repo root automatically. Add a line so it scans without being asked every time:

   # CLAUDE.md
   Before opening a pull request that touches .tf files, call
   audytx scan_terraform on the full set of Terraform files in the
   repo — not just the diff; cross-resource reasoning needs the unchanged
   files too. Fix every High and Critical finding, then re-scan. For the
   mechanical ones, autofix_terraform applies the sound fixes and
   re-scans in one pass.

   Pass the complete file set, not only the changed files — audytx builds IAM trust graphs, attack paths, and DLQ-identity edges across resources, and it can't reason about a file you withhold.

The gaps a single-file review can't see

These are the patterns our study found frontier models ship by default — the reason this setup earns its place:

• Secrets Manager rotation. The model creates an aws_secretsmanager_secret for a database password but never the companion aws_secretsmanager_secret_rotation. A pattern matcher flags the secret for a missing attribute; audytx finds the absent companion resource and fires with a rationale that names it.
• IAM privilege-escalation paths. A role that can iam:PassRole into a compute service it also controls is an escalation even when every individual statement looks scoped. audytx traces the documented paths across the trust graph.
• Async wiring without alarms. The model reaches for dead-letter queues but forgets to alarm on them; it enables encryption but omits rotation. Reliability and observability gaps dominate the count; outright security holes are rare.

The suppressions matter just as much — the false positives audytx removes so the agent doesn't chase them: a DLQ that isn't asked to have its own DLQ, a Lambda flagged for a missing DLQ when nothing invokes it asynchronously, an IAM role that reads privileged but can't escalate. Each suppression is printed with its reasoning, so the agent (and you) can check the call.

Same engine on every other surface

The MCP loop is the pre-PR half. For the review half, install the GitHub App — it posts one comment per pull request with the same findings, inline annotations, and SARIF to GitHub Code Scanning, so anything the agent missed still gets caught on the PR. Building CloudFormation instead of Terraform? The same agent loop works with scan_cloudformation — same rule ids, same context.

Common questions