Menu
19 official servers · 33 platforms checked · Last updated
Compliance MCP servers: which GRC platforms have one, and what can it change?
As of 12 Oct 2026, 19 of the 33 compliance and GRC platforms we have checked document their own MCP server. LogicGate and Strike Graph are read-only, and 10 others offer a read-only connection; Anecdotes, Scrut Automation, SimpleRisk and Thoropass limit writes to one area; 8 platforms document delete access; and only 7 say that assistant actions are logged.
Asking whether an MCP server is secure, or what your SOC 2 auditor will test? See SOC 2 controls for AI agents and MCP servers.
Platforms with an official MCP server for compliance data are the rows; the rest we have checked are named under the table. Payment plays no part in who is listed or in what order.
Compliance platforms' MCP servers compared
One row per platform with an official MCP server, from each vendor's own documentation. Access shows the most an assistant can do; on most platforms the connecting user's role narrows it further. Not documented means the vendor does not say, not that the feature is missing.
| Official MCP server | Access | Who can connect | Read-only and revocation | AI actions logged | Setup | Checked | |
|---|---|---|---|---|---|---|---|
| Anecdotes MCP docs | Official | Limited write: it can update existing control details; everything else is read-only | Whoever holds an API key or token | Read-only: not documented; revoke the key or token | Not documented | Hosted by the vendor; API key; Claude, Cursor, VS Code / Copilot, any MCP client; Requires the AI suite package and AI consent opted into on the account | 12 Oct 2026 |
| Carbide MCP docs | Official | Read and write | Whoever holds an API key or token; the vendor turns it on | Can connect read-only; revoke the key or token | Not documented | Hosted by the vendor; API key; Claude, ChatGPT, any MCP client; Available on all Carbide plan tiers; the customer's advisor enables it on request | 12 Oct 2026 |
| CISO Assistant MCP docs | Official | Read, write and delete | Any user, limited to their role | Can connect read-only; revocation not documented | Each call runs as the token's own user, and the docs say a shared token would stop the audit trail telling callers apart. No MCP-specific log is described | Runs locally or self-hosted; API key; Claude Desktop, Claude Code, ChatGPT, Microsoft Copilot Studio, LM Studio, OpenWebUI, Cursor; Community and Pro, SaaS or on-premises; no separate fee stated. The audit log is Pro only | 13 Oct 2026 |
| Comp AI MCP docs | Official | Read, write and delete | Any user, limited to their role; no setup step | No read-only option; revoke one connection | Not documented | Hosted by the vendor; OAuth; Claude, Claude Code, ChatGPT, Cursor, VS Code / Copilot, any MCP client; plans not stated | 12 Oct 2026 |
| Conveyor MCP docs | Official | Read, write and delete | Any user, limited to their role; no setup step | Can connect read-only; revoke one connection | User sessions and tool calls are logged; the questions asked and the answers returned are not | Hosted by the vendor; OAuth or API key; Claude, ChatGPT, Cursor, any MCP client; plans not stated | 12 Oct 2026 |
| Drata MCP docs | Official | Read, write and delete | Any user, limited to their role; an admin turns it on first | Can connect read-only; revoke one connection | Drata's AI product page says it keeps full audit logging; its setup docs do not say where MCP actions are recorded | Hosted by the vendor; OAuth; Claude, ChatGPT, Cursor, Microsoft Copilot (Copilot Studio); plans not stated | 12 Oct 2026 |
| LogicGate MCP docs | Official | Read-only | Certain roles only; an admin turns it on first | Read-only server; revoke one connection | Not documented | Hosted by the vendor; OAuth or API key; Claude, Claude Code, Microsoft Copilot Studio, ChatGPT, Cursor; Included with Risk Cloud API access at no extra cost; each user needs the MCP Access role entitlement | 12 Oct 2026 |
| Oneleet MCP docs | Official | Read, write and delete | Whoever holds an API key or token; no setup step | Can connect read-only; revoke the key or token | Each change is attributed to the signed-in user or the service key, and the audit log can be filtered by either; the docs do not say whether MCP actions are marked as such | Hosted by the vendor; service token or OAuth; Claude, Claude Code, ChatGPT, Codex, Cursor, Devin, any MCP client; plans not stated | 12 Oct 2026 |
| Optro | Official | Not documented | Not documented | Not documented | Optro says AI actions are logged and traceable; it does not document where or in what detail | Hosting not stated; sign-in not stated; any MCP client; plans not stated | 12 Oct 2026 |
| SafeBase by Drata MCP docs | Official, early access | Read and write | Any user, limited to their role; an admin turns it on first | Can connect read-only; revocation not documented | Not documented | Not documented | 12 Oct 2026 |
| Screenata MCP docs | Official | Read and write | Not documented | Not documented | Each MCP request is recorded in the workspace's API audit log with the tool that ran, alongside CLI and API traffic. | Hosted by the vendor; OAuth; Claude Code, Cursor, Windsurf, OpenAI Codex, any MCP client; Included in every plan; a free assessment account gets assessment permissions only until the workspace is activated | 12 Oct 2026 |
| Scrut Automation MCP docs | Official | Limited write: it can upload a file and attach it, or a link, to an existing evidence item; nothing can be edited or deleted | Any user, limited to their role; no setup step | Read-only: not documented; revoke one connection | Logged like the user's own actions (for example, an upload by that user); the AI client itself is not recorded | Hosted by the vendor; OAuth; Claude, ChatGPT, Cursor, Claude Code, Codex, any MCP client; All plans: Foundation, Growth and Scale | 12 Oct 2026 |
| Secureframe MCP docs | Official | Read, write and delete | Certain roles only | Not documented | Not documented | Hosted by the vendor; OAuth or API key; Claude Code, Claude Desktop, Cursor, any MCP client; plans not stated | 12 Oct 2026 |
| SimpleRisk MCP docs | Official | Limited write: it can queue AI-drafted control tests for a person to approve; everything else is read-only | Whoever holds an API key or token; an admin turns it on first | Can connect read-only; revoke the key or token | Not documented | Self-hosted; API key; Any MCP client (Streamable HTTP); Needs the API Extra; the control-test proposal tool also needs the AI Extra. Core is free | 13 Oct 2026 |
| Sprinto MCP docs | Official | Read and write | Certain roles only; an admin turns it on first | Can connect read-only; organisation-wide switch only | Not documented | Hosted by the vendor; OAuth or API key; Claude, Claude Code, ChatGPT, Cursor, Microsoft Copilot Studio, any MCP client; plans not stated | 12 Oct 2026 |
| Strike Graph MCP docs | Official | Read-only | Certain roles only; the vendor turns it on | Read-only server; revoke one connection | Not documented | Hosted by the vendor; OAuth; Claude, ChatGPT, any MCP client; Scale and Enterprise plans, or with AI Consumption credits; unlimited use once enabled | 12 Oct 2026 |
| Thoropass MCP docs | Official | Limited write: evidence requests only (submit, attach files, comment, and delete its own comments and attachments); audits, controls and policies are read-only | Not documented | Can connect read-only; revoke one connection | Not documented | Hosted by the vendor; OAuth; Claude, Claude Code, any MCP client; All customers (per GA blog); access arranged through the customer success manager | 12 Oct 2026 |
| Vanta MCP docs | Official, public preview (its docs say beta) | Read, write and delete | Any user, limited to their role | Can connect read-only; revoke one connection | Not documented | Hosted by the vendor; OAuth; Claude Code, Claude Cowork, Cursor, Perplexity, Codex, ChatGPT, any remote MCP client with OAuth; Essentials, Plus, Professional and Enterprise plans; some areas return data only where that module is enabled on the account | 12 Oct 2026 |
| Whistic MCP docs | Official, beta | Read, write and delete | Whoever holds an API key or token | Read-only: not documented; revoke the key or token | Not documented | Hosted by the vendor; OAuth or API key; Claude, ChatGPT, Cursor, any MCP client; plans not stated | 12 Oct 2026 |
No official MCP server for compliance data documented as of 12 Oct 2026: ComplyJet, Delve, Hyperproof, Iru, ISMS.online, OneTrust Certification Automation, Onspring, RealCISO, Scytale, ServiceNow Integrated Risk Management, TrustCloud, Trustero and Zania.
Whether Apptega runs an official MCP server is not established; we found no public documentation.
From vendor documentation, help centres and release notes on the date in the Checked column. We describe what each vendor documents; we have not connected or tested these servers.
Which compliance platforms have an official MCP server?
Of the 33 platforms we have checked, 19 document an official MCP server: Anecdotes, Carbide, CISO Assistant, Comp AI, Conveyor, Drata, LogicGate, Oneleet, Optro, SafeBase by Drata, Screenata, Scrut Automation, Secureframe, SimpleRisk, Sprinto, Strike Graph, Thoropass, Vanta and Whistic. An official server is one the vendor documents and supports itself. Connectors that third parties build on a vendor's public API, such as Truto, StackOne or Zapier, are not official servers, and they show up in search results even for vendors with their own server, including Sprinto and Drata.
A server for another product does not count. Iru's server reaches its device management product, with actions such as lock and erase, and not compliance records. Hyperproof and OneTrust run servers that search their API documentation rather than your data, and ServiceNow's MCP console documents no risk or compliance tools.
Stage words vary and are not comparable. Vanta calls its hosted server a public preview in release posts and a beta in its documentation. Drata launched its server as "experimental" in June 2025; its current documentation carries no stage label. Whistic labels its server beta and SafeBase calls its server early access. The table shows each vendor's own word where it uses one, and Official where it uses none.
Vanta and Secureframe both replaced an early local server with a hosted one in 2026. Vanta archived its open-source server on 17 June 2026 and Secureframe archived its read-only beta on 17 July 2026, so articles from 2025 describe software that no longer ships.
What can an AI assistant read, change or delete through these servers?
Compliance MCP servers differ most in write access. LogicGate and Strike Graph are read-only. Anecdotes, Scrut Automation, SimpleRisk and Thoropass limit writes to one area. Carbide, SafeBase by Drata, Screenata and Sprinto document writes but not deletes. CISO Assistant, Comp AI, Conveyor, Drata, Oneleet, Secureframe, Vanta and Whistic document delete actions. Optro does not document what its server can change.
Within one area, Anecdotes can update existing control details, Scrut can attach files to existing evidence, SimpleRisk can queue AI-drafted control tests for a person to approve, and Thoropass keeps its writes to evidence requests.
What a delete covers differs by platform. Drata's server can delete evidence, risks and vendors but not controls. Vanta's deletes reach privacy records and risk registers, while its tests and controls stay read-only. Secureframe's deletes cover comments, knowledge-base items and some system security plan records. Sprinto documents no tool named delete, though it can archive and close records. CISO Assistant's deletes reach only vulnerabilities and task templates.
On most platforms the connecting user's role caps what the assistant can do. Where a service key or API token is used instead (Anecdotes, Carbide, Conveyor, Oneleet and Whistic all document one), the key's own grants apply, so the key needs the same review as any other credential. Drata adds a second limit: an administrator chooses the scopes, and the assistant gets only what both the scopes and the user's role allow.
Can you connect an assistant to your compliance platform read-only?
Of the 19 official compliance MCP servers, 2 are read-only for every connection: LogicGate and Strike Graph. Carbide, CISO Assistant, Conveyor, Drata, Oneleet, SafeBase by Drata, SimpleRisk, Sprinto, Thoropass and Vanta let you connect an assistant read-only. Comp AI documents no read-only option, so a connection carries the user's own rights. Anecdotes, Optro, Screenata, Scrut Automation, Secureframe and Whistic do not document whether a connection can be limited to reading.
How a read-only connection is made differs by vendor: a read-only scope, key, role or server setting. Ask which one applies to the plan you are buying, and confirm it blocks writes before you connect an account that can write.
Who can connect an assistant, and how is access revoked?
Who may connect an assistant to a compliance MCP server varies across the 19 official servers. CISO Assistant, Comp AI, Conveyor, Drata, SafeBase by Drata, Scrut Automation and Vanta let any user connect, limited to their role. LogicGate, Secureframe, Sprinto and Strike Graph limit connection to certain roles. Anecdotes, Carbide, Oneleet, SimpleRisk and Whistic let whoever holds an API key or token connect. Optro, Screenata and Thoropass do not document who can connect. Drata, LogicGate, SafeBase by Drata, SimpleRisk and Sprinto need an admin to turn the server on first. Carbide and Strike Graph have the vendor turn it on for you.
Revoking access is documented less often than granting it. Comp AI, Conveyor, Drata, LogicGate, Scrut Automation, Strike Graph, Thoropass and Vanta document revoking a single connection. On Anecdotes, Carbide, Oneleet, SimpleRisk and Whistic, access ends when the key or token is revoked. Sprinto documents only an organisation-wide switch. CISO Assistant, Optro, SafeBase by Drata, Screenata and Secureframe do not document how access is revoked. Ask any vendor how to revoke one person's connection before you approve the first one, and how long an assistant's access lasts before it must sign in again.
Are AI actions logged where your auditor can see them?
Of the 19 official compliance MCP servers, 7 have a vendor statement that assistant actions are logged: CISO Assistant, Conveyor, Drata, Oneleet, Optro, Screenata and Scrut Automation. The detail varies. Scrut records actions taken through its server like the user's own actions, though not which AI client made them. Oneleet attributes each change to the signed-in user or service key. Conveyor logs sessions and tool calls but not the questions and answers. Screenata records each MCP request in its API audit log with the tool that ran. CISO Assistant runs each call as the token's owner, so its audit log, a Pro feature, attributes changes to that user. Drata's AI product page says it keeps full audit logging, without saying where, and Optro says AI actions are logged and traceable, without saying where or in what detail.
Anecdotes, Carbide, Comp AI, LogicGate, SafeBase by Drata, Secureframe, SimpleRisk, Sprinto, Strike Graph, Thoropass, Vanta and Whistic do not document whether assistant actions are logged. Carbide says administrators can review what its built-in assistant did; it does not say whether calls from your own assistant are logged the same way.
At audit time, an auditor testing logical access and monitoring will want to see what a non-human identity did. If your platform cannot show which changes came from an assistant, keep that record on the assistant or gateway side.
What does connecting an MCP server change in your SOC 2 scope?
An assistant that can write to your compliance platform is a privileged integration, so it falls under the same criteria as any other: who connected it and with what rights, monitoring of what it did, change control over its configuration, and vendor management for the AI provider that processes the data. No AI-specific Trust Services Criteria exist yet; auditors apply the existing ones.
On an MCP connection you choose the assistant, and its provider becomes a vendor that processes your compliance data. Controls, evidence and risk records leave the platform for Claude, ChatGPT or whichever client you connect, under that provider's terms rather than the platform's. Check its data retention and training terms and add it to your vendor inventory before you connect, as you would for any other vendor with access to that data.
Frequently asked.
What is an MCP server in a compliance platform?
MCP (Model Context Protocol) is an open standard for connecting AI assistants to software. A compliance platform's MCP server lets an assistant such as Claude, ChatGPT or Cursor query controls, evidence and risks, and on most platforms change them, with the permissions of the person or key that connected it.
Is an MCP server secure enough for compliance data?
Its security depends mostly on the access you grant. Prefer per-user sign-in over shared API keys, grant the narrowest access the work needs, confirm how to revoke a connection, and find out whether the platform logs what the assistant did. The protocol itself is not certified; your vendor and your configuration are what an auditor examines.