Przejdź do głównej treści
Public AI connector guide

Connect your AI assistant to Kernion safely

Use Hermes, Claude Code, or any MCP-compatible client to inspect, draft, QA, and operate Site Studio through scoped connector presets — without handing an agent a platform-admin session.

Sign in to issue a connector See how it works

Agency partners issue from their agency Integrations page. Platform support issues tenant-limited support connectors from the agency cockpit.

Quickstart

Three safe steps, no public secrets

The public guide teaches the connection model. Real credential creation, one-time secrets, revoke/rotate, and scope selection stay behind authenticated role-specific screens.

1

Choose the narrow preset

Start with Draft editor for everyday site work. Publisher is explicit because it changes what visitors see.

2

Issue a scoped credential

Sign in, open the correct connector settings page, choose the tenant or agency scope, and copy the one-time bearer value into your secret store.

3

Configure the client with an env var

Point your MCP client at the /mcp endpoint and reference the bearer value through an environment variable, not inline config.

Presets

Use presets instead of hand-picking dangerous checkboxes

Authenticated screens still show advanced capability details. The normal path is preset-first: draft-only by default, explicit publish power only when needed.

Read + QA

Inspect selected sites, schemas, and QA findings without editing drafts.

Risk posture

Safest posture for audits, support triage, and tool discovery.

Capabilities

  • studio.read
  • studio.schema.read
  • studio.qa.run

Draft editor

Recommended first

Create and repair Site Studio drafts while keeping visitor-visible publishing off.

Risk posture

Recommended default. Humans still review and publish live changes.

Capabilities

  • studio.read
  • studio.schema.read
  • studio.draft.blocks.write
  • studio.draft.content.write
  • studio.qa.run

Designer

Draft content plus theme, media, and navigation changes for selected sites.

Risk posture

Can reshape the visitor journey. Review carefully before publishing.

Capabilities

  • studio.read
  • studio.schema.read
  • studio.draft.blocks.write
  • studio.draft.content.write
  • studio.qa.run
  • studio.theme.write
  • studio.media.write
  • studio.navigation.write

Publisher

Draft editing plus explicit permission to publish public Site Studio changes.

Risk posture

Visitor-visible. Use short TTLs, one operator per credential, and audit every key.

Capabilities

  • studio.read
  • studio.schema.read
  • studio.draft.blocks.write
  • studio.draft.content.write
  • studio.qa.run
  • studio.publish

Agency operator

Agency workspace operations such as client, app, and readiness management.

Risk posture

Can affect client onboarding and readiness workflows. Keep tenant allowlists tight.

Capabilities

  • studio.read
  • studio.schema.read
  • studio.draft.blocks.write
  • studio.draft.content.write
  • studio.qa.run
  • apps.read
  • apps.manage
  • readiness.read
  • readiness.manage
  • clients.view
  • clients.manage
Publisher rule: Publishing changes what visitors see. Do not grant Publisher unless the operator explicitly owns public launch responsibility for that site.

Client setup

Use the live domain and an environment variable

After issuing a credential, store the bearer value as AGENCY_CONNECTOR_MCP_BEARER. These examples deliberately show only the variable reference.

Add this under mcp_servers. Store the bearer value in the process environment or an env file loaded by Hermes.

# ~/.hermes/config.yaml
mcp_servers:
  kernion_site:
    transport: http
    url: https://agency.kernion.io/mcp
    headers:
      Authorization: "Bearer ${AGENCY_CONNECTOR_MCP_BEARER}"

Security model

  • No keys are issued on this public page. It teaches the flow only.
  • Raw bearer keys are shown only once on authenticated issuance screens.
  • Capabilities are stored with the credential and re-checked against the live role on every call.
  • Draft writes stay private until someone with explicit publish power publishes them.
  • Every tool call is rate-limited, audited, and tied to the credential that made it.

Troubleshooting

Common connector failures are intentional guardrails

Most setup failures are scope, credential, or transport mismatches. Fix the boundary rather than broadening the connector blindly.

401

Invalid, missing, revoked, or expired token

Re-copy the environment variable from your secret store, rotate the connector if the raw key was lost, and confirm the Authorization header uses Bearer plus the variable value.

403

Capability or tenant scope denied

Use a higher preset only when the operator truly needs it, or add the exact tenant site to the credential allowlist.

405

Wrong transport method

The MCP endpoint is Streamable HTTP over POST. GET is intentionally not a server-sent event stream.

Ready to connect an agent?

Start with Draft editor, verify tools/list, then escalate only when the person requesting it owns the risk.

Sign in See plans