<!-- mcp-name: io.github.infraveilhq/agent-guard -->

infraveil-guard

A seatbelt for your AI agent. Put a governed, tamper-evident gate in front of the destructive things an agent can do — rm -rf, DROP TABLE, terraform destroy, git push --force, kubectl delete namespace, DELETE FROM … with no WHERE. The agent proposes; the dangerous ones are blocked until a human approves them out of band; every decision is written to a local hash-chained ledger you can verify.

Offline by design: no account, no network, no telemetry. It runs entirely on your machine. Open your network tab — it talks to nobody.

pip install infraveil-guard

---

Why

Coding agents (Claude Code, Cursor, and friends) are great until the one time they run rm -rf in the wrong directory, or drop the production database to "fix a migration." You don't want to read every command — you want the catastrophic ones to stop and wait for you. That's all this does, and it does it well.

Wire it into your agent

Add it as an MCP server. For Claude Code / Cursor / any MCP client:

{
  "mcpServers": {
    "infraveil-guard": {
      "command": "infraveil-guard"
    }
  }
}

Then add one rule to your agent's instructions (CLAUDE.md, system prompt, etc.):

Before running any shell command, SQL statement, or infrastructure/cloud operation, first call guard_action with the exact command. Only proceed if it returns proceed: true. If it returns decision: "blocked", stop and ask me to approve it — I'll give you a one-time code to pass back as approval_code.

That's it. Safe commands sail through (and are logged). Dangerous ones stop.

How approval works (the part that matters)

When the agent hits something dangerous, guard_action returns blocked and an action_id. The agent cannot approve itself — by construction, not by good behavior. You approve in your own terminal:

$ infraveil-guard approvals
1 action(s) blocked, waiting for approval:

  [9b58e9c499b3]  CRITICAL  CRITICAL risk: drop table (+0 more). Irreversible.
            DROP TABLE users;
            approve with:  infraveil-guard approve 9b58e9c499b3

$ infraveil-guard approve 9b58e9c499b3

  Action requesting approval
  id:        9b58e9c499b3
  risk:      CRITICAL  (IRREVERSIBLE)
  why:       CRITICAL risk: drop table. Irreversible.
  command:
    DROP TABLE users;

  Approve this action? [y/N] y

  APPROVED. Give the agent this one-time code:

      8f2510

  It is valid for 15 minutes and works exactly once.

You hand the agent 8f2510; it calls guard_action("DROP TABLE users;", approval_code="8f2510"); the guard checks it, lets it through once, and records the approval. The code is minted only by the human CLI, is single-use, and expires — so an agent can't forge or replay it.

Inspect everything — trust nothing

Every decision (allowed, blocked, approved, denied) is appended to a hash-chained ledger at ~/.infraveil-guard/ledger.jsonl. Editing, deleting, reordering, or inserting any line breaks the chain:

$ infraveil-guard verify
{ "ok": true, "count": 42, "message": "Hash chain verified across 42 entries - no tampering." }

$ infraveil-guard log 10        # the last 10 decisions, raw

It's ~400 lines of plain stdlib Python. Read it. That's the point.

Tools (MCP)

Tool	What it does
guard_action(action, approval_code="")	Gate an action before running it. Returns proceed true/false.
assess_action(action)	Classify blast radius without recording or gating.
verify_ledger()	Verify the tamper-evident ledger's hash chain.
recent_decisions(limit=20)	The most recent decisions, newest first.

Configuration

Env var	Default	Meaning
INFRAVEIL_GUARD_THRESHOLD	high	Gate actions at/above this severity: none|low|medium|high|critical.
INFRAVEIL_GUARD_MODE	enforce	enforce blocks dangerous actions; audit logs everything but never blocks (use it to watch your agent before you trust the gate).
INFRAVEIL_GUARD_HOME	~/.infraveil-guard	Where the ledger and approval queue live.

What this is — and isn't

It is a high-signal classifier + an out-of-band human-approval gate + a tamper-evident local log. It's the smallest honest version of "a human approves before anything irreversible happens."

It is not a sandbox. It works because your agent is told to route actions through guard_action — a cooperative guardrail, not an unbypassable jail. That is a deliberate trade: in exchange you get something you can install in one line, read end to end in an afternoon, and run with no account, no network, and no dependency on anyone else's infrastructure — including ours. Nothing here calls home, checks a license, or needs a server to keep working. It does one job and owns it: stop the catastrophic actions and wait for a human. Yours to fork and run forever.

Related

Part of Infraveil's open-source tools for running backends you own:

• infraveil-lockin — scan a repo and score how locked in your app is to one cloud provider (npx infraveil-lockin).
• infraveil — a policy DSL you enforce in CI, plus offline audit-ledger and signature verification.

License

AGPL-3.0-or-later. Use it, fork it, read every line. If you run a modified version as a network service, share your changes. © Infraveil Corporation.