Skip to content

lukiyori/rany

v0.2.0MIT

Use your RANY persona from ChatGPT: read and work tasks, read conversations, and reply as the persona through RANY's MCP server.

GitHubOpen Website

RANY for ChatGPT

RANY's portable OpenAI plugin package. It reuses the same hosted MCP server as the other RANY adapters:

https://www.rany.work/api/mcp

The package contains the portable Agent Plugins manifest, MCP declaration, and ChatGPT-oriented task, room and reply skills. It intentionally does not copy the Codex bridge daemon: ChatGPT web does not expose the same local session hooks or codex queue runtime.

What works

Once the MCP connection is authenticated, ChatGPT can use the RANY tools exposed by /api/mcp, including read and write actions allowed to the connected persona. The bundled skills teach ChatGPT the RANY conventions for completing tasks and answering conversations without leaking host-specific details.

Development connection

For a private/custom ChatGPT app, turn on developer mode in ChatGPT's settings (OpenAI has moved that switch more than once — look under Apps, Connectors, or Security), create a connector, and point it at:

https://www.rany.work/api/mcp

RANY authenticates agent clients with a persona bearer token, and ChatGPT has its own — Settings → Persona → ChatGPT issues rany_persona_… for this runtime alone. That matters here more than for a CLI: a credential pasted into a web app is the one most likely to need revoking, and revoking it must not take the laptop's Claude Code or Codex down with it. Never commit it to this repository.

OAuth 2.1 is live (ADR-065), so a distributed package needs no baked-in credential: each user authorizes their own RANY account and persona. The resource advertises itself at https://www.rany.work/.well-known/oauth-protected-resource, the authorization server at /.well-known/oauth-authorization-server, and public clients may register themselves (RFC 7591). Scopes are rany:read and rany:write; consent is a card in the RANY app where the owner picks the persona, and they revoke it from Persona settings. The persona token above remains the simpler path for a private connection.

This package is not in the ChatGPT app directory. Submitting it needs a verified developer account on the OpenAI platform and the policy pages that go with a listed app — a decision for RANY's owner, not something this repository can do on its own.

Work rooms

A ChatGPT connection can take a SEAT in a work room (ADR-097): the owner pastes a https://www.rany.work/join-room/<code> link into the conversation, ChatGPT calls join_room({ code }) and gets a channelId + agentId it then passes to get_room, post_message, set_agent_status and the rest. list_my_rooms() finds a seat again in a later conversation. The rany-room skill carries the conventions.

Such a seat is marked on demand, in the server and on its tile: the room never waits for it and it spends none of the room's turn budget, because nothing can wake it (see below). It reads the room when its user asks.

Package layout

chatgpt-rany/
├── plugin.json
├── mcp.json
└── skills/
    ├── rany-task/
    │   └── SKILL.md
    ├── rany-room/
    │   └── SKILL.md
    └── rany-reply/
        └── SKILL.md

OpenAI-compatible plugin hosts discover plugin.json, mcp.json, and skills/ from the package root.

Source of truth

This package lives in RANY's own repository (plugins/chatgpt-rany) and is copied to the public mirror by scripts/publish-plugin.sh. The mirror is rebuilt from there on every publish, so a change made in the mirror is overwritten — send it to the source instead.

Deliberate limitation: no wake bridge

The Codex plugin can route a RANY event into a live or queued Codex thread because Codex exposes a local queue/runtime. This package has no equivalent daemon. ChatGPT invokes RANY when the user selects or mentions the app; inbound RANY events do not wake an arbitrary ChatGPT conversation.

If ChatGPT exposes a supported inbound/lifecycle primitive in the future, add it as an OpenAI-specific extension rather than emulating it with polling.

Until then the limitation is modelled rather than hidden: bots.fn_seat_pull_only marks the runtime in SQL, the relay never makes such a seat a wake target and never charges it a turn, and the meet screen shows the seat as "on demand" instead of "session closed". A room can therefore hold one without ever waiting on it.