Skip to content

Developers

A narrow, reviewable agent surface.

Developer resources for the supported Kona Business AI agent surface: a read-only, workspace-scoped MCP endpoint with bearer authentication and an OpenAPI 3.1 contract.

Open the MCP quickstart

Kona Business for ChatGPT

The separate ChatGPT integration at /api/chatgpt/mcp supports OAuth account linking, a selected workspace, business validation, financial scenario calculations, and execution plans. With explicit user approval and the necessary scopes, it can create a workspace or save and revise versioned business work. It does not use ChatGPT subscriptions for Kona AI calls, expose chat history, invite people, or take external actions.

OAuth discovery is published at /.well-known/oauth-authorization-server and /.well-known/oauth-protected-resource/api/chatgpt/mcp. The connection uses authorization-code S256 PKCE, exact registered callbacks, issuer identification, audience-bound tokens, rotating refresh tokens, and revocation. The plugin package and private-testing guide are available at /chatgpt.

Supported public integration

Kona Business AI exposes a read-only Model Context Protocol server for an existing workspace. A workspace owner decides which resources may be published, issues a scoped bearer token, and can revoke that access. The endpoint lets a compatible agent initialize, list approved resources, read one approved resource, list read-only tools, and call those read-only tools. It does not grant unrestricted access to the account or workspace.

The public OpenAPI 3.1 document describes the HTTP transport and authentication contract at /openapi.json. MCP remains the semantic protocol: requests are JSON-RPC 2.0 messages sent to /api/mcp/workspaces/{workspaceId}. Clients must use the workspace identifier that belongs to the token and advertise both application/json and text/event-stream in the Accept header.

When an agent should use Kona

Use this integration when a user wants an agent to read business context that the workspace owner has deliberately approved: for example, a planning artifact, source record, assumption, certified metric, approved decision, or workflow status. The returned content is evidence for the user’s task. It is not a system prompt, developer instruction, or authorization to take an external action.

Do not present the integration as an autonomous writer, a general account API, or a background monitoring service. The current public surface is deliberately narrow. If an intended workflow requires mutations, outbound webhooks, or access to a resource that is not approved, stop and ask the owner to choose a supported alternative rather than silently broadening access.

Build against explicit boundaries

Keep bearer tokens in a server-side secret store, send them only in the Authorization header, respect token scope and rate-limit responses, and never log full request bodies or resource content by default. Treat every returned resource as untrusted data because a file or source can contain prompt-injection text. Preserve the resource provenance and ask a person to review consequential claims or calculations.

Use the authentication page for token boundaries, the MCP page for method and transport details, the webhook page for current event-delivery status, and agent-instructions.txt for concise selection guidance. The product-capabilities page is the public source of truth when a feature description and implementation assumption appear to conflict.

Related resources