Connect AI Agents
Kantis includes an MCP (Model Context Protocol) server at https://mcp.getkantis.com/. Connect your AI assistant to it and the assistant can read your organisation’s compliance data — policies, tasks, Context Hub, frameworks, risks, vendors, people (including who accepted which policies, security awareness training and device checks), integrations, the Statement of Applicability, evidence documents, audit findings, security questionnaires, Trust Center access requests and the audit trail — and answer questions like “which policies are still in draft?” or “what is left before our Stage 1 audit?”. For ISO 27001 it also gets a requirement-by-requirement view: each clause and Annex A control with its Statement of Applicability row, the policies that cover it and the evidence behind it.
Your external auditor can connect their own assistant in the same way, with read-only auditor access; see For Auditors.
Trust Center access requests contain personal data of people outside your organisation (names, emails, companies and job titles of those who asked for your documents), so they are a separate permission on the Kantis authorisation page. Grant it only if you are comfortable sending that data to the AI assistant you use; connections approved before 24 September 2026 need to reconnect to include it.
Access is read-only by default. The assistant cannot change anything in Kantis unless your organisation opts in to AI suggestions (see below), and even then it can only propose draft policy versions for a person to review.
Suggested policy changes (optional)
Section titled “Suggested policy changes (optional)”Assistants can suggest improvements to a policy as a new draft version. The assistant sends only the passages it wants to change (“replace this with that”), and Kantis applies them to the policy’s existing text, so the rest of the policy and its formatting stay exactly as they were. Each change is highlighted, with a short comment showing what it was, what it is now and why. Nothing is published: the policy owner gets an email, reviews the draft in the policy’s Versions tab, resolves every comment, and then submits it through the normal approval workflow.
If a passage the assistant wants to change can’t be found exactly once in the policy, Kantis saves nothing and tells the assistant which changes to fix, so a half-applied suggestion never reaches your team.
Suggestions are off by default. To use them:
- An organisation owner or admin turns on Allow AI agents to suggest changes in Settings → AI Agents.
- You reconnect your assistant and approve the separate “make suggestions” permission on the Kantis authorisation page. Existing connections stay read-only.
Suggested versions are marked AI in the versions list and attributed to the person who connected the assistant. Limits apply: at most 3 open AI suggestions per policy and 20 per organisation per day. If someone edits the policy after the assistant read it, the suggestion is refused and the assistant is asked to re-read the policy first. If the policy already has an unpublished draft (someone’s edits in progress, or a version awaiting approval), the assistant must ask you whether to build on that draft or to ignore it and edit the published version; Kantis refuses the suggestion until you’ve chosen, and the owner sees which choice was made.
Evidence file contents (optional)
Section titled “Evidence file contents (optional)”Assistants can also read what is inside your evidence files: the text of attached documents (PDF, Word, and Excel as CSV) and screenshots as images. Kantis reads the file and passes on its contents; it never hands out a download link. This covers files attached to tasks and files uploaded in evidence forms.
File reading is off by default. An organisation owner or admin turns on Let AI agents read evidence files in Settings → AI Agents. It applies to every connection in the organisation, including your auditor’s, and needs no reconnect. Each file read is recorded in the audit trail with the person, the file and the AI tool. Turning it off stops reads straight away.
Connect from Claude
Section titled “Connect from Claude”- In Claude, open Settings → Connectors → Add custom connector.
- Enter
https://mcp.getkantis.com/and add the connector. - Claude opens a Kantis authorisation page. Sign in to Kantis if asked, check the listed access, and click Allow access.
- Ask Claude something about your organisation to confirm it works.
The connection is bound to the organisation that was active in Kantis when you approved it. To connect a different organisation, switch organisations in Kantis first, then reconnect.
Connect from ChatGPT
Section titled “Connect from ChatGPT”- In ChatGPT, open Settings → Connectors and add a custom connector (available on plans that support MCP connectors).
- Enter
https://mcp.getkantis.com/as the server URL. - Approve access on the Kantis authorisation page, as above.
Other MCP clients
Section titled “Other MCP clients”Any MCP-capable client can connect to https://mcp.getkantis.com/ using OAuth — client registration happens automatically. If your tool asks for endpoints, they are discoverable at https://mcp.getkantis.com/.well-known/oauth-authorization-server.
Managing and revoking access
Section titled “Managing and revoking access”Every connection appears in Kantis under Settings → AI Agents, showing who connected which agent, what it can read, when it was last used and when it expires. You can disconnect your own agents at any time; organisation admins and owners can disconnect anyone’s. Disconnection takes effect immediately.
Connections expire 30 days after approval. To keep using an assistant, simply approve access again.
How access is secured
Section titled “How access is secured”In plain words: the assistant acts as you, sees only what you can see, changes nothing unless your organisation opts in to reviewed suggestions, and everything it does is recorded.
The specifics:
- Your own login. Access is granted through OAuth 2.1 with PKCE, bound to your Kantis sign-in and your active organisation at the moment of approval. There are no shared credentials, and Kantis never shows your password to the assistant.
- Your own permissions, re-checked every time. Every request the assistant makes is checked against your live organisation membership and role, at the time of the request. If your access changes or your account is deactivated, the assistant’s access changes with it, immediately.
- Read-only by design, enforced server-side. A connection only carries read scopes unless you explicitly approve suggestions, and the service identity behind it holds no permission to change, publish, approve or delete anything. Its single extra ability, creating a draft policy version for human review, is off until an admin enables it. Read-only does not depend on the assistant behaving.
- Strict tenant isolation. A connection is bound to one organisation. Records from any other organisation are not visible, even if their identifiers are known.
- Hashed, expiring credentials. Kantis stores only cryptographic hashes of access tokens, never the tokens themselves. Access tokens live 15 minutes; connections expire after 30 days and any suspicious token reuse revokes the whole connection automatically.
- Everything is audited. Connections, the data access they enable, every evidence file an assistant reads, and every revocation are recorded in your organisation’s audit trail with the user, the agent and the action.
For your auditor
Section titled “For your auditor”The Settings → AI Agents page is designed to serve as third-party access-review evidence (ISO 27001:2022 A.5.18 / SOC 2 CC6.x): it lists every AI connection with its holder, scope, last use and expiry, and supports immediate revocation by administrators. Revocations are written to the audit trail. Reviewing this page as part of your periodic access review, and exporting or screenshotting it as evidence, is the intended workflow. If your auditor wants to connect their own AI tool, see For Auditors. An auditor’s connection is read-only and leaves out your sales questionnaires and Trust Center access requests.