← Kapālins for ClaudeDOCUMENTATION · CONNECTOR FOR CLAUDE
The Kapālins connector for Claude
Ask Claude where your organisation stands on AI security, what is happening right now, and what your policy would do before you act. Fourteen read-only tools, signed in as you, with your role and your plan applied on every call.
Overview
What it does
The connector lets Claude read your organisation’s Kapālins data the way you would in the console, and nothing more. Claude calls it as you: it sees only your organisation, only what your role may see, and only what your plan includes. Every figure it returns comes from the same calculation as the console page it names, with a link to that page.
It is read-only. It cannot change a policy, close an incident, approve a request or switch anything on or off; those decisions stay in Kapālins. It never returns who did something: people and devices are counted, never named.
Tools
Every tool, in three groups
Claude chooses the tool from your question. Each one belongs to an ability you approve when you connect.
Where you stand
Abilities: read posture · read compliance
| Tool | What it answers |
|---|---|
Get the AI Trust Index and what is lowering itget_trust_index | Your score and band, the pillars pulling it down, as on the AI Trust Index page. |
Show what Kapālins can and can’t seeget_coverage_gaps | Each way AI is used in your organisation (browser, coding agents, gateway, workspace, devices) and how each is covered. |
List what to fix firstlist_recommendations | Open recommendations, ranked by impact, each with the page that fixes it. |
Get compliance readiness by frameworkget_compliance_readiness | Readiness, evidence and open gaps for each framework you track. |
What’s happening
Abilities: read incidents · read integrity · read the AI found in use · read data security · read AI cost · read approvals
| Tool | What it answers |
|---|---|
List open AI incidentslist_open_incidents | Open security, compliance and cost incidents, most severe first, with counts across the whole queue and whether each is overdue. |
Get an AI incidentget_incident | One incident’s severity, status, what raised it and when, with signal summaries masked. |
List AI applications and their integritylist_ai_applications | Every AI application with its integrity state and the kind of its weakest link. |
Explain an AI application’s integrity ratingget_application_integrity | Why an application is rated as it is, across shape, identity, reach, supply chain, behaviour and authority, each marked measured, not applicable or not measured. |
List AI tools found in uselist_discovered_ai | AI tools found on devices and through your identity provider, approved or not, with counts of people and devices. |
List agents, MCP servers and skills on deviceslist_device_agents | Coding agents, MCP servers and skills found on managed devices, what each can reach, and whether its hooks are governed. |
Get sensitive data that reached AIget_data_exposure | Which kinds of sensitive data reached AI in the last 7 days, and how much was stopped or redacted. Types and counts only. |
Get AI usage, cost and budgetsget_ai_spend | Requests, cost and budgets by team, application and model. Prices appear only where your console shows them. |
List requests waiting for a decisionlist_pending_approvals | Held agent actions and policy-exception requests waiting for an approver. Claude cannot approve or deny them. |
Before you act
Ability: check against policy
| Tool | What it answers |
|---|---|
Check text or an action against my organisation’s AI policycheck_against_policy | What your policy would do with this text or tool call at the AI gateway, in the browser and in coding agents: allowed, coached, redacted, held for approval or blocked, and the rule that decides. Nothing is sent or enforced, and the text is not stored. |
Before you connect
What you need
In Kapālins
- A Kapālins organisation, and an account in it.
- An Owner or Admin has turned the connector on (it is off until someone does).
- Your own password set, if you were invited with a temporary one.
- A member’s account. Platform operators and partner staff cannot connect Claude to a customer’s organisation.
In Claude
- A Claude plan that lets you add custom connectors.
- Your own Claude account. Each person connects as themselves; there is no shared key.
Setup
Four steps, one of them in Claude
- Turn it on for your organisationAn Owner or Admin opens My Integrations › AI Providers › Assistant connectors › Claude connector and turns on People can connect Claude. The same page shows the connector URL.
- Add it in ClaudeIn Claude, open Customize › Connectors › Add › Add custom connector, name it Kapālins and paste the connector URL. Leave sign-in on Sign in now and the OAuth client on Use Claude’s published identity.https://mcp.kapalins.com/mcpIn Claude Code:claude mcp add --transport http kapalins https://mcp.kapalins.com/mcp
- Tell it your organisationClaude opens a Kapālins page that asks for your organisation’s address, for example yourcompany.app.kapalins.com. It then sends you to your own sign-in, with your single sign-on and second factor.
- Approve read accessThe page lists each ability under Where you stand, What’s happening and Before you act, and says if your role or plan leaves one out, with the reason. Select Allow read access, and you are back in Claude.
Access
Who can use what
Each ability follows the console permission and plan feature behind it. If you cannot see a page in the console, Claude cannot read it either. Roles and plans are checked again on every call, so a change takes effect at once.
| Ability | Roles | Plan feature |
|---|---|---|
| Read posture | Every member. Coverage also needs Owner, Admin, Security Lead or Compliance Manager. | AI Trust Index (every plan) |
| Read compliance | Owner, Admin, Compliance Manager, Analyst or Viewer | Compliance |
| Read incidents | Owner, Admin, Security Lead or Compliance Manager | Every plan |
| Read integrity | Every member who can see AI Systems | Autonomous Integrity |
| Read the AI found in use | Owner, Admin, Security Lead or Compliance Manager | AI Estate Discovery |
| Read data security | Owner, Admin, Security Lead or Compliance Manager | Data Security |
| Read AI cost | Owner, Admin, FinOps Manager or Analyst | AI Cost Observability |
| Read approvals | Owner, Admin or Security Lead | Held agent actions: every plan. Policy-exception requests: AI Gateway & Prompt Firewall |
| Check against policy | Every member | AI Gateway & Prompt Firewall (every plan) |
Roles that use only the Governed AI Workspace cannot use the connector. It is not offered to an organisation that uses its provider’s own brand with no “Powered by” line; My Integrations says so instead of showing an address.
Examples
Ask it the way you would ask your security team
“Brief me for the board: where are we on AI risk this week?”Trust Index, compliance readiness, open incidents, sensitive data stopped, AI found in use and agent activity, in one answer.
“What’s our AI Trust Index, and what is pulling it down?”The score and band, the weakest pillars and the first fixes.
“What should we fix first?”The ranked recommendations, each with the page that fixes it.
“What are my open critical AI incidents, and which are overdue?”Open incidents, most severe first, with a link to each.
“Which AI tools are people using that we haven’t approved?”Discovered tools with counts of people and devices, never names.
“How ready are we for ISO 42001?”Readiness, evidence and open gaps for that framework.
“Would our policy let me paste this customer list into ChatGPT?”The outcome at each enforcement point and the rule that decides. Nothing is sent.
Data and privacy
What Claude sends, what we keep
Sent and recorded
- Claude sends only each tool’s arguments, such as an incident reference or a framework name.
- For each call we record the tool, its outcome and how long it took, in your organisation’s activity trail.
- For a policy check we record the length of the text and a one-way fingerprint of it.
Never received, kept or returned
- Your Claude conversations, memory or files.
- The text you ask us to check.
- Prompt text from incidents, or who sent it.
- Names of people or devices. They are counted, never named.
Text that came from outside your organisation, such as an application’s name, is marked as untrusted in each answer so Claude treats it as data, not instructions. Read section 3.5a of our privacy policy.
Security
The details a security team asks for
| Sign-in | OAuth 2.0 authorization code with PKCE (S256). Claude identifies itself with its client metadata document; registration accepts only Claude’s own callback or a loopback address. Each person signs in to their own organisation, with its single sign-on and second factor. |
| Tokens | Opaque and bound to the connector alone; only their SHA-256 is stored. Access tokens last an hour. Refresh tokens rotate on every use and lapse after 30 days unused; one presented twice ends the connection. |
| Every call | The organisation comes from the connection, never from a tool argument. The organisation’s switch, its status, your membership and your role are read again on every call. |
| Read-only | Every tool is marked read-only. Writes are refused below the tools as well, so a fault cannot change your data. |
| Limits | 120 calls a minute per connection. Answers are capped in size; Claude is told to ask for fewer rows if one is too large. |
| Where it runs | Google Cloud, Sydney (australia-southeast1), behind Cloudflare. |
Troubleshooting
If something doesn’t work
| What you see | Why, and what to do |
|---|---|
| “Your organisation has turned the Claude connector off.” | An Owner or Admin turns it on in My Integrations. Turning it off ends every connection. |
| “No organisation signs in at that address.” | Use the address you sign in to the console at, for example yourcompany.app.kapalins.com, not your email. |
| Claude says Connection expired and offers Reconnect | The connection lapsed after 30 days unused, an administrator ended it, or you chose Reconnect Claude to add abilities. Reconnect and approve again. |
| A tool says the ability was not approved | You connected before that ability was available to you. On the Claude connector page, select Reconnect Claude (or remove the connector in Claude and add it again), then approve it. |
| “This sign-in request has expired” | The sign-in page is valid for 10 minutes. Go back to Claude and connect again. |
| “Needs a role that can …” | Your role does not include that page in the console. The message names the roles that do; ask an administrator. |
| “… is not in your organisation’s plan.” | The plan feature behind that tool is not included. Your administrator or provider can tell you more. |
| “Sign in to your organisation’s console and set your own password first.” | Your account still has a temporary password. Sign in once to the console and set your own. |
| A figure says no data or not measured | There is nothing to measure there yet, so the answer says so rather than showing zero. |
| “This could not be confirmed right now. Try again shortly.” | A check could not be completed. Nothing was returned in its place. Try again; if it persists, contact support. |
Disconnect
Ending a connection
In Claude, remove the connector from Customize › Connectors. An Owner or Admin sees every connection in the organisation and can end any of them on My Integrations › AI Providers › Assistant connectors › Claude connector. Removing a member or turning the connector off ends their connections at once.
Support
Help and security reports
Product questions: [email protected]. Security concerns or vulnerability reports: [email protected], read by the security team.
The connector is built by Kapālins. Claude is a product of Anthropic. This connector is an integration, not an Anthropic product.