GxP Gateway Audit (mcp-glp-audit-proxy) - Setup and Verifica

    1

    Set up GxP Gateway Audit: hash-chained audit trail for MCP and HTTP(S) APIs, gateways and LLMs. Policy, body_mode, mTLS, verify and regulatory-report.

    Free

    1 useSecurity scanned

    Works with the AI Agents you already use

    Claude CodeClaude CodeCursorCursorCodex CLICodex CLIDeepSeekDeepSeekMuseMuse+22 more

    GxP Gateway Audit (mcp-glp-audit-proxy) - Setup and Verifica

    Example session with this skill installed

    An approver must sign off before a write goes to the Excel server. Walk me through the human approval flow in mcp-glp-audit-proxy.

    • Read your context and instructions
    • Compiled the gxp gateway audit

    Human approval applies when require_approval_for_writes is true and the tool matches write_tools.

    Approval flow

    1. The operator calls the write tool as usual (with a reason). The proxy does not forward it. It logs a pending_approval entry and replies with approval_id=<uuid> args_hash=<sha256>. The approval expires after approval_ttl_seconds (default 3600).
    2. A different person, running the proxy with GLP_PRINCIPAL_ROLE=approver, calls glp_approve. Input schema: approval_id (string, required) and meaning (required, one of approved, reviewed, rejected):
    {"approval_id": "3f0c9a2e-7b1d-4c55-9e1a-2d6f8b0c4e11", "meaning": "approved"}
    
    1. The operator repeats the exact same call (same tool, identical arguments) and adds the id in _meta:
    {
      "name": "excel__write_cell",
      "arguments": {"cell": "B2", "value": 42, "_glp_reason": "Correct transcription error"},
      "_meta": {"io.github.antonio-castellon/approvalId": "3f0c9a2e-7b1d-4c55-9e1a-2d6f8b0c4e11"}
    }
    

    The call is forwarded only if all checks pass; otherwise it is refused and logged as approval refused (<problem>):

    • unknown: no such approval id.
    • self: the requester cannot approve their own write (glp_approve also refuses this directly).
    • role: the decision was not made by the approver role.
    • meaning: only approved authorises. reviewed is recorded but does not authorise the write.
    • rejected: the approver rejected it.
    • mismatch: tool name or SHA-256 of the canonical {tool, arguments} differs from the request. Any change to arguments needs a new approval.
    • expired: past expires_at.

    Next steps

    • Anchor the log head outside the machine: mcp-glp-audit-proxy-verify data/audit.jsonl --print-head prints one stdout line: seq=<n> entry_hash=<hash>.

    Connects securely to your tools. The creator never sees your data.

    About this skill

    GxP Gateway Audit (mcp-glp-audit-proxy) is a free, open-source (MIT) local proxy that audits MCP tool calls AND HTTP(S) APIs, gateways and LLM endpoints. It applies an allow/deny policy and writes every mediated call, including denials and failures, to an append-only, RFC 8785 canonical, SHA-256 hash-chained JSONL audit trail. It is designed to support GLP / 21 CFR Part 11 / EU Annex 11 workflows. Data stays on your infrastructure; there is no phone-home.

    Version 0.4.0 adds HTTPS as a second inbound/outbound channel. Six flows are supported: mcp2mcp, mcp2api, api2mcp, api2api, mcp2llm and api2llm. With no listen block it stays a stdio MCP proxy; with listen it serves HTTPS on TLS 1.3 only, and your client simply changes its base URL.

    This skill is the setup and validation companion. What your agent learns:

    • Install the proxy (Python 3.11+) and wire it into Cursor, Claude Desktop or any stdio MCP host, or put it in front of an HTTP(S) API or LLM endpoint.
    • Choose audit.body_mode per your QA decision: hash_only (default), redacted or full. The SHA-256 of every body is always stored, for MCP arguments/results and HTTP bodies alike.
    • Configure HTTPS client identity: mtls, bearer or mtls_and_bearer. There is no unauthenticated mode. Token files are checked for 0600 on POSIX and for a restricted NTFS ACL on Windows.
    • Write a deny-by-default policy for MCP tools and for HTTP routes (METHOD /path), with required reasons and optional second-person approval for writes.
    • Understand fail-closed behaviour: if the record cannot be written, the result is not returned; a broken hash chain stops the process and writes a critical alert naming where it broke.
    • Run regulatory-report: an offline report on your config (the service is not started) that lists, per regulation and country, which settings are supported and which parameter to change.
    • Archive with archive-manifest and verify --manifest, close segments with reason and approver, and use review for audit trail review with exception report and CSV export.
    • Anchor the chain head off the machine with --print-head and generate the OQ-style verification report as evidence for your own validation.

    A paid Pro edition adds signed checkpoints, WORM storage, PDF/CSV reports, an IQ/OQ pack and SSO/OAuth.

    The skill and the proxy produce evidence that supports your own validation, SOPs and reviews. They do not certify anything. Tamper-evident is not tamper-proof; the approval record is not an electronic signature.

    Repo: https://github.com/antonio-castellon/mcp-glp-audit-proxy PyPI: https://pypi.org/project/mcp-glp-audit-proxy/ License (server): MIT. Contact: antonio@castellon.ch

    Find more skills like this on the Claude Skills Marketplace, the Codex Skills Marketplace and the Gemini Skills Marketplace, or browse Claude Skills for LLM Ops.

    Detects but does not prevent edits or deletion of the audit file (tamper-evident, not tamper-proof). Requires Python 3.11+. The approval record is not an electronic signature. Arguments are stored in full in the audit trail. Identity depends on host environment variables.

    How to use

    1. 1

      Connect your agent to the Agensi MCP

      Copy the prompt and paste it into Claude, ChatGPT, Cursor or any agent that supports MCP. It sets up the connection for you. Manual setup

    2. 2

      Sign in to Agensi

      Your agent shows a sign-in link. Sign in and click Allow.

    3. 3

      Ask for what you need

      Your agent loads this skill and uses it on your task. It always asks before buying a skill.

    Reviews

    No reviews yet

    Be one of the first to try it. Every listed skill goes through the trust checks below.

    Security scanned

    Checked by our 8-point scan before listing

    1 use

    Times used to date

    Free to use

    No account required to browse

    Trust & safety

    Security scan

    Security scan passed 29 days ago

    • Free with an Agensi account

    Software
    Python 3.11+
    pip
    oruv (Python tool runner)
    Glp Audit Proxy MCPStdio MCP host

    mcp-glp-audit-proxy is on PyPI. Any stdio MCP host works, such as Cursor or Claude Desktop.

    Permissions

    Terminal / Shell
    Read Files
    Write Files
    Network Access

    Network access is only used to install mcp-glp-audit-proxy from PyPI, and for the proxy to reach upstream MCP servers that you configure yourself (HTTPS / TLS 1.3). The skill itself needs no network access. Shell runs pip/uv and the verify CLI; file access is the local config and audit log.

    Listed29 days ago
    Updated23 days ago

    What's inside

    Frequently Asked Questions

    No software is certified on its own. The proxy is designed to support GLP / 21 CFR Part 11 / EU GMP Annex 11 workflows. Fitness depends on your validated use, SOPs and reviews; the OQ-style report is evidence for your own validation.

    Yes. mcp-glp-audit-proxy Community is MIT on GitHub and PyPI. This skill is the paid setup and verification guide.

    No. The proxy runs locally and writes the audit file on your infrastructure. There is no phone-home.

    Edits, deletions and reordering are detected, not prevented. Coordinated truncation of log and head file is caught by comparing with the head line you stored outside the machine.

    No. It records that a second person approved a specific write, but it is not a 21 CFR Part 11 electronic signature.

    The proxy fails closed: the call is not completed, later calls are blocked until restart, and the skill walks you through the reconciliation SOP.

    antonio@castellon.ch

    New in LLM Ops

    Free