---
title: "Auditable agentic compliance"
canonical: "https://auditbadger.com/auditable-agentic-compliance/"
last-updated: "2026-09-04"
---

# Auditable agentic compliance

Auditable agentic compliance means an AI agent can do the work, a human approves every change, and the audit trail proves which was which. Agent-native, human-gated GRC.

Point Claude Code, Codex, Cursor, or any MCP client that can send a bearer token at an AuditBadger account. It can read controls, draft policies, chase evidence gaps and prepare vendor reviews. It cannot change anything: every mutation arrives as a proposal, a person approves it, and a separate step applies it, with the agent named as the actor and the approving human named next to it.

## What we claim, and what we do not

We are not the first compliance platform with an MCP server. Vanta, Drata and Secureframe shipped theirs before we shipped ours. The narrower thing we believe is actually new:

> As far as we can tell, AuditBadger is the first compliance platform where an external AI agent can do real compliance work under a mandatory human approval gate that cannot be switched off, with every applied change attributed to the agent in a retained audit trail.

- **Mandatory**: not a setting or a default. There is deliberately no flag that restores direct agent writes, and a test fails the build if someone adds one.
- **Cannot be switched off**: including by us, including by an admin. The proposal step is architecture, not configuration.
- **Attributed and retained**: filter the audit log by actor channel `agent` for every agent-caused change, its proposal, and its human approver. Confirmations are stored separately from conversation content.

If someone shipped this before us, email hello@auditbadger.com and we will correct this page.

## Why the audit trail is the point

- SOC 2 expects privileged actions to be attributable to an accountable individual, not an autonomous agent or a generic API key.
- ISO 27001 Annex A logging expects a record of what happened, when, and who was responsible.
- The EU AI Act adds record-keeping and human oversight obligations for automated tools acting inside processes a person remains accountable for.
- A direct agent write is a change with no request, no approval and no accountable person. A human-approved proposal is an ordinary change management record: request, approval, application, timestamps.

## Mechanics

| | |
|---|---|
| Endpoint | `https://{account}.auditbadger.com/agent/mcp` (Streamable HTTP, bearer agent token) |
| Tools | 118: 58 reads with observed versions, 49 propose operations, lifecycle, upload, and deprecated-alias tools |
| Scopes | Per module, read or read-and-propose; an agent never sees more than the person who issued its token |
| Judgment | Risk assessments, implementation status, policy publication, check acceptances: decided one at a time, never bulk-approved |
| Expiry | Undecided proposals expire on their own |
| Uploads | PNG, JPEG, GIF, WebP, PDF; active-content PDFs refused; image metadata stripped; recorded `not_scanned` (no malware scanner today) |

Not autopilot, and not continuous monitoring: automated checks run monthly, plus on demand.

Included in the subscription at $250/month flat. No per-user fees, no per-agent pricing.

## Useful links

- Technical tour: [/headless-compliance](/headless-compliance)
- Feature page: [/features/api-and-mcp](/features/api-and-mcp)
- Developer hub: [/developers](/developers)
- Credentials: [/auth.md](/auth.md)
- Pricing: [/pricing](/pricing)