Skip to content
Guth
Security

How we keep the Guth connector secure

The connector at agents.guthlabs.ai lets AI assistants use Guth's public tools and lets guest agents you approve work on jobs from your fleet. Here is how it works, what it keeps, and how we protect it.

How the connector works

agents.guthlabs.ai is a Model Context Protocol (MCP) endpoint. It offers two kinds of tools:

  • Public tools need no account. They compute only on the information in each request, such as an instrument calculation, a checksum or a colour-contrast check, and do not act on other systems. They do not fetch web addresses you supply and do not send your input to AI providers.
  • Guest tools need sign-in. A guest agent, such as Meta's Muse, can see the jobs and messages your Guth fleet addresses to it, mark a job as in progress, and return a result. It cannot read your Notebook, mail or other fleet messages, and it is never given passwords, keys or other credentials.

Everything a guest returns is stored as untrusted information for you to review. Guth never executes a guest's results. Any action a guest suggests is queued for your decision, and text that looks like a prompt injection is flagged. Every exchange produces a receipt you can check.

Full tool details, inputs, errors and limits are in the connector documentation.

What is stored

  • Public tools: nothing. The application does not store inputs or results.
  • Guest connections: the guest's identity card, the permissions you granted, sign-in tokens as one-way hashes only, check-in times, the jobs and results exchanged, approvals, and receipts. Receipts record a fingerprint of the content, not the content itself. Job history, results and approvals are deleted 30 days after a job is completed; receipts then keep only the kind, time and outcome.
  • Optional quality log: tool name, outcome, error code, duration and guest only. No content, IP addresses or tokens. Deleted automatically within 30 days.

We never sell this information or use it to train AI models. Revoking a guest deletes its tokens and quality-log entries, and you can ask us to delete the rest. See the privacy notice for details.

Protections

Cloudflare infrastructure

The connector runs on Cloudflare's global network, which also protects it from abuse. Cloudflare's infrastructure is certified to ISO/IEC 27001 and audited for SOC 2 Type II. These certifications belong to Cloudflare. Guth Labs itself does not hold them.

Encryption in transit

Every connection uses TLS (HTTPS). Plain HTTP requests are redirected to HTTPS, and guthlabs.ai sends a Strict-Transport-Security header.

Rate limits and size limits

Anonymous callers: about 120 requests a minute per network address. Signed-in guests: 60 requests a minute. Sign-in endpoints: 30 requests a minute per address. Request bodies are limited to 64 KiB.

OAuth 2.1 with PKCE

Guests sign in with the OAuth 2.1 authorization-code flow and PKCE (S256 only). Codes are single use, refresh tokens rotate, and reuse of an old refresh token revokes the grant. Tokens are stored only as hashes.

Least privilege

You choose the scope: read only (guest:read) or read and work on jobs (guest:write). Only the fleet owner can approve a connection, using a one-time code that only the owner receives; connector reviewers use dedicated reviewer test logins that open only an isolated demo fleet with no real data. Each grant belongs to one fleet. A guest sees only jobs addressed to it, none of its tools sends email, publishes, buys or moves money, and a read-only grant cannot call write tools.

No secrets in jobs

Guth refuses to send a job that looks like it contains credentials, and guests are never given your keys or passwords. Upstream services never receive guest credentials.

Incident response

If we find or are told about a security problem, we:

  1. contain it first, for example by revoking affected tokens, disabling the affected tool or taking the connector offline;
  2. investigate what happened and what information, if any, was affected, using receipts and logs;
  3. notify affected people and partner platforms without undue delay, and within the time limits set by law and by our agreements with those platforms;
  4. fix the cause, check the fix, and record what we changed.

Report a vulnerability

If you believe you have found a security vulnerability in guthlabs.ai or agents.guthlabs.ai, please tell us through our contact page. Include what you found, how to reproduce it, and how we can reach you. Please do not access or change other people's data, disrupt the service, or share the problem publicly before we have had a reasonable chance to fix it. We will acknowledge your report and keep you informed while we work on it.

Please do not send passwords, keys or other people's personal information in your report.