Infraveil Agent Guard
- Updated June 22, 2026
- infraveilhq/infraveil-guard
Infraveil Agent Guard is an offline MCP server that gates destructive agent actions behind out-of-band human approval and records tamper-evident local audit decisions.
About
Infraveil Agent Guard is an offline MCP server that sits in front of an AI agent's destructive actions and blocks high-severity commands until a human approves them out of band, logging every decision in a tamper-evident local ledger. Developers reach for it when coding agents can run shell, SQL, or cloud operations and you want a cooperative seatbelt rather than blind trust. Add infraveil-guard as an MCP server, instruct agents to call guard_action before risky operations, and approve blocked action IDs from your own terminal.
- Offline MCP server with guard_action, assess_action, verify_ledger, and recent_decisions tools
- Blocks rm -rf, DROP TABLE, terraform destroy, git push --force, and similar high-risk actions by default
- Human approval via separate terminal; agents cannot self-approve by design
- No account, network, or telemetry required; AGPL open source
Infraveil Agent Guard by the numbers
- Data as of Jul 7, 2026 (Skillselion catalog sync)
claude mcp add --env INFRAVEIL_GUARD_THRESHOLD=YOUR_INFRAVEIL_GUARD_THRESHOLD --env INFRAVEIL_GUARD_MODE=YOUR_INFRAVEIL_GUARD_MODE --env INFRAVEIL_GUARD_HOME=YOUR_INFRAVEIL_GUARD_HOME infraveil-guard -- uvx infraveil-guardAdd your badge
Show developers this MCP server is listed on Skillselion. Paste this into your README.
| Package | infraveil-guard |
|---|---|
| Transport | STDIO |
| Auth | None |
| Last updated | June 22, 2026 |
| Repository | infraveilhq/infraveil-guard ↗ |
How do I stop an AI agent from running catastrophic shell, SQL, or infrastructure commands without a human explicitly approving them?
Block destructive agent shell, SQL, and infra commands until a human approves out-of-band, with a tamper-evident local ledger.
Who is it for?
Developers using Claude Code, Cursor, or other MCP clients who want a local approval gate and verifiable audit log for destructive agent actions.
Skip if: Teams needing a mandatory unbypassable fleet-wide sandbox with centralized policy; that requires the full Infraveil control plane instead.
What you get
Returns proceed true or blocked with an action_id, requires human approval codes for gated actions, and appends verifiable hash-chained ledger entries.
Recommended MCP Servers
1Claw VaultHSM-backed vault secrets for AI agents (JIT fetch) plus prompt-injection and threat scanning.2
1passwordMCP server for 1Password service accounts — tools and resources for vaults and credentials18
402sentinel McpAssess an x402 counterparty's risk BEFORE paying: allow/review/block, scored on-chain (Base).1
A2adependencyinspector McpRemote MCP for A2A dependency inspector MCP, structured receipts, audit logs, and reviewer-ready evi
A2a Governance Bridge McpA2A Governance Bridge MCP server.A2aidentitytoll McpRemote MCP for A2A caller identity, scope policy, verdict receipts, and audit history.FAQ
Can the agent approve its own blocked actions?
No. Dangerous actions require out-of-band human approval in a separate terminal; the agent only receives proceed false until approved.
Does Agent Guard need network access?
No. It runs fully offline with no account, telemetry, or external dependency after installation.
What MCP tools does it expose?
guard_action to gate actions, assess_action to classify risk, verify_ledger to check the hash chain, and recent_decisions for audit review.