OSuite Docs

ChatGPT Apps Directory

Publish OSuite into ChatGPT as a remote MCP app with OAuth, review tools, and buyer-readable audit links.

Use this page when

Read this page when you want OSuite to appear as a public ChatGPT app instead of a local runtime package.

What this path is

This path packages OSuite as a remote MCP app for ChatGPT. It keeps the governance contract, but presents a smaller and more productized tool surface to end users.

How it differs from the Codex plugin

  • ChatGPT Apps Directory is the remote MCP app path for ChatGPT
  • Codex Plugin is the packaged rollout path for Codex app, CLI, and IDE

Choose this path when you want public discovery, OAuth-based install, and a tool surface that reads like a business app.

What OSuite exposes

The current app exposes five product-level tools:

  • review_planned_work
  • check_if_approved
  • summarize_runtime_exposure
  • record_final_outcome
  • open_audit_record

Session setup, decision capture, replay packaging, and session closing stay inside OSuite.

What OSuite captures automatically on this path

This lane is stronger than a free-text audit note, but weaker than a hook runtime such as Claude Code.

  • review_planned_work automatically opens the runtime session, creates the governed action, records the OSuite preflight decision, and stores the declared decision context.
  • review_planned_work is not the business action itself. It writes a bounded OSuite governance record and returns whether the connected client should continue, wait, or stop.
  • blocked work can be closed directly from the preflight path without waiting for a second tool call.
  • summarize_runtime_exposure returns a read-only summary of runtime security posture, approval-bound actions, blocked actions, top blast-radius agents, and exposure backlog.
  • record_final_outcome now closes the governed runtime session by default and marks the closure as connector-declared evidence unless a stronger observation basis is supplied.
  • if the connector never sends a final outcome, OSuite can still reconcile a stale remote session into a visible auto-closed record.

What this path still does not do

This ChatGPT Apps lane does not claim black-box capture of arbitrary internal ChatGPT tool calls. The evidence produced here is connector-declared and OSuite-governed rather than full runtime-observed execution telemetry.

The ChatGPT-facing response is intentionally minimal. OSuite returns the action ID, approval state, replay link, verification link, exposure summary, and outcome state needed for the review, but not raw internal action payloads, auth tokens, full logs, or unrelated user profile data.

Transport and auth

  • transport: remote MCP over streamable HTTP
  • auth: OAuth 2.0 authorization code flow with PKCE
  • install channel: ChatGPT Apps Directory

Before submission

Confirm these items before review:

  • the remote MCP endpoint is reachable
  • OAuth metadata is published
  • the app can bind the user to a workspace
  • the tool list is stable
  • audit links resolve back into OSuite

Test flow

  1. Create an app that points to https://studio.osuite.ai/api/connectors/osuite/mcp.
  2. Complete the OSuite OAuth flow.
  3. Confirm tools/list succeeds.
  4. Run review_planned_work.
  5. If the action escalates, run check_if_approved.
  6. Run summarize_runtime_exposure when you want the current runtime security posture and exposure backlog.
  7. Close the path with record_final_outcome. The governed runtime session closes automatically unless you explicitly keep it open.
  8. Open the audit record and confirm the replay view shows the capture profile as connector-declared / OSuite-preflight.

What success looks like

The app is ready when a user can review work, complete the OAuth flow, and open an audit-ready record without seeing internal runtime vocabulary.

Next pages

Cookie settings

Essential cookies keep sign-in and workspace routing stable. You can separately choose Analytics cookies and Preference cookies.