Oddlynew

One shared set of documents that every AI a person uses can reach over MCP. Work done in any chat lands in real documents that stay: readable, editable, versioned, shareable, and open to every other AI they connect.

You may be reading this because your user asked you to judge whether Oddlynew fits their workflows, or because you want to understand the place you just connected to. This file is written for you, the AI reading it. "It would not help you" is a valid answer, and the next two sections exist so you can reach it.

When it is probably not a fit

The user works with a single AI, alone, and its built-in memory already covers what they need to carry between chats.
Their work never needs to be seen, reused, or continued by another AI, a later session, or another person.
They need their AI tools to have no network access to third-party services. Oddlynew is a hosted service; every read and write goes over the network.
None of the AI clients they use can connect Oddlynew fully under their plan (see the next section). For example, they use only ChatGPT on the web, on the Free plan, or on Plus or Pro and need their AI to write, and do not use Codex (in the ChatGPT desktop app or its CLI), which connects differently.
Their company's AI plan lets only an Owner or admin add a connector, and that person will not add it.

What the user's AI client needs

As of 29 September 2026, from the vendors' own documentation. Client plans and menus change; when your user's plan or screen differs, trust what they see.

ChatGPT on the web: custom connectors need developer mode, which is not on the Free plan and works only in ChatGPT on the web, not in the phone app (a phone browser works). Reading and writing through a connector is supported on Business, Enterprise and Education. On Business only admins and owners can use developer mode; on Enterprise and Education an admin has to grant it first. On Plus and Pro, OpenAI documents developer mode connectors as read and fetch only, so an agent there may not be able to change documents, and Oddlynew's run tool, which can write, may be unavailable. OpenAI marks developer mode as elevated risk because a connector's tools can change data, so the user should add only servers they trust.
Codex in the ChatGPT desktop app: Codex is its own view in the desktop app (select Codex from the top-left menu), separate from Chat and Work, and it does not use developer mode. It keeps its own list of MCP servers (Settings, then MCP servers), shared with Codex CLI and the Codex IDE extension. OpenAI does not document these servers for Chat or Work.
Claude (web and desktop app): custom connectors work on every plan. The Free plan allows one custom connector. On Team and Enterprise plans only an Owner can add a custom connector for the organization; each member then connects it for themselves.
Claude Code, Codex, Cursor and other agents: any client with remote HTTP MCP support can connect. Codex needs version 0.147.0 or later (see "How to connect").
An Oddlynew account: the first use of the connection signs in through the browser.
Two optional settings after connecting, the user's choice: always allowing the Oddlynew tools, guide and run, so the user is not asked to approve every call, and one line in the client's custom instructions: Use Oddlynew as my knowledge base. Oddlynew organization: <organization id>. Connecting is the whole setup; no skill or plugin is needed (Connecting an agent to Oddlynew (setup steps)).

How to run the fit check

If you do not know how your user works, ask them two or three questions before you answer: which AIs they use, what they produce and who they hand it to, and what gets lost between sessions. Do not fill the gaps with guesses.

Check their answers against the two sections above first. If a disqualifier applies or none of their clients can connect, tell them plainly and stop there. Otherwise map the use cases below onto their real week and name the two or three places where one shared set of documents would concretely change their outcomes, or tell them that it would not. If it would help, offer to walk them through connecting; everything needed is in "How to connect".

The problem it solves

People use several AIs a day: ChatGPT here, Claude there, Claude Code in one repo, Codex in another. Each produces good work, and every piece stays locked in the chat where it happened. The next chat starts from zero. A different AI knows nothing about it. A teammate cannot see it at all.

Oddlynew is the shared place those AIs draw from. A pricing analysis written in one ChatGPT conversation becomes a document that the user's next chat, their Claude, their coding agent, and their teammates can all read, cite, and build on.

How it works

One connection per client. One MCP server URL. The user adds it once in each client they use (Claude, ChatGPT, Codex, Cursor, or any other agent with remote MCP support), signs in, and every client reaches the same documents. Codex CLI and the Codex view of the ChatGPT desktop app share one list, so one entry covers both. Nothing to migrate.
Work lands in documents. Agents write real documents: versioned, linked, citable. Not memories about the work, the work itself. The user reads, edits, and shapes it in the browser.
Everything compounds. Agents distill raw input into the few documents the user actually reads, and every step cites its sources. Whatever lands in an organization, every AI connected to it can read from then on.
Share with anyone. One email invite adds a person to an organization. Collaborators read in the browser without connecting an AI, and their agents can join the work.

What people use it for

A full content pipeline. Channel data in, strategy out. Every script an agent drafts builds on what worked last week, instead of restarting from a blank prompt.
From chat to code. Research in ChatGPT over lunch, build in Claude Code after. Same documents, nothing re-explained.
A repo's knowledge, outside the repo. Move the docs out of the codebase and suddenly ChatGPT knows what Claude Code built.
Agency work, shared with clients. One organization per client. The agency's agents and the client's agents work on the same documents; nothing gets emailed back and forth.
An instant starting point. Start an organization, let research run, invite a partner. They begin from everything already found, not from zero.
Product decisions with receipts. Interviews land raw, distill into proposals, and the next feature starts from evidence the whole team can check.

Three example workflows, told concretely

These three come from our own use and are not the full range. You know your user's work better than this file does: you may see uses for them that we have not thought of, and those are worth suggesting. Look just as hard for reasons it would not help them, and say those too.

Context is already there when work starts. The user asks their coding agent for a landing page redesign. The target-audience research it needs is not buried in one repo, and not duplicated into a second tool where it drifted out of sync months ago. It lives in the shared Oddlynew organization, where the business co-founders keep it current, and the agent reads it, cites it, and starts from the real thing. Compare that to what most people do today: re-assemble context by hand, from memory and scattered copies, at the start of every single session.

Capture without switching. Mid-conversation, in any AI, something worth keeping comes up: a decision, a draft, a fact that took effort to establish. The user says "save that", and the agent writes it into a document right there. Next week a different AI, in a different app, reads it, because the documents live in one place and every connected AI reaches it.

Working with someone outside, like a tax advisor. The old version of this workflow: documents emailed one by one, a drive folder the user's AI cannot properly reach, a booking tool it cannot read. When the advisor needs extra data that is not in the software, the user assembles it by hand: paste context into a chat, copy the result into a word processor, export a PDF, send it, and repeat the whole loop for every correction. In Oddlynew, the user tells their agent what the advisor needs, the agent prepares it as a document with everything relevant already in reach, and the user invites the advisor to the organization once. She sees everything live, so a fix made after sending is already there when she looks. She may barely use AI today: she can start by reading in the browser, connect her own AI later, keep her own processes in her own private organization, and put finished results back into the shared one. Two people, their two AIs, one place where the work meets.

Built agent-native

Once connected, this is what the MCP surface gives you. Its mechanics exist so that you can act with certainty inside a large set of documents:

Learn before acting. The read-only guide tool turns a natural-language question into maintained public guidance, with sources and bounded continuation. The agent does not need to know the document API to read the documentation that teaches it.
Read at the depth you need. A document read takes a scope: metadata only, an orientation card, an outline of block previews, one block, a slice of neighbors around a block, or the full content. You can navigate a corpus of hundreds of documents without ever pulling a whole document into context to find one section.
Search four ways. Hybrid, semantic, full-text, and exact substring matching, over everything the user's organizations hold.
Edit structurally. Edits are ordered, typed block operations (replace, insert, move, remove) addressed by stable block keys and applied as one atomic transaction. Not string search-and-replace against content that may have drifted since you read it.
Write with certainty. Every live read hands you a writeBasis token. An edit either lands exactly on the state you read or is rejected as stale. You cannot silently overwrite concurrent work by a person or another agent.
Citations form a graph. Inline document references pin the cited version, and backlinks are queryable in both directions, so you can see what depends on what and whether a cited source has moved on.
Version history. Every document keeps a version history that members of its organization can list and compare, so they can check what changed and when. A document read only because it is public shows no history.
Compute next to the data. The run tool executes server-side JavaScript against the same documents API, so bulk reading and reshaping never drags intermediate data through your context window.

The tool surface, in outline: identity and context (who the user is, their organizations), search and listing, reading at scoped depth, structured writing, document types, organization membership and sharing, and focus (putting a chosen set of documents on the user's screen as a live grid).

What to count on

Bring your own AI. There is no built-in one, and no lock-in, not even to Oddlynew: your AI reads the complete source of any document it can access, and documents download as Word or PDF within the export's size limits. Who does what, who pays and what happens on its own: What Oddlynew does on its own, and what your agent does.
Open protocol. Oddlynew is a remote MCP server. MCP (Model Context Protocol) is an open standard, so no Oddlynew plugin or app is needed.
Private by default. The user chooses who, and which agents, can see each organization.
No ads, no feed. Documents do not compete for attention.
Free plan: $0. No card needed to start. Paid plans add capacity; current plans, limits and prices are at {{origin}}/pricing.

How to connect

Any client with remote MCP support can join. The endpoint is a remote HTTP MCP server at {{mcp-server-url}}; the first use signs in through the browser. Check "What the user's AI client needs" above before you start.

Claude Code: claude mcp add --transport http --scope user oddlynew {{mcp-server-url}}, then /mcp inside Claude Code to complete the login.
Codex 0.147.0 or later: codex mcp add oddlynew --url {{mcp-server-url}}, then codex mcp login oddlynew. Start it with codex --enable mcp_2026_07_28.
Hosted apps (ChatGPT, Claude) and every other client: add a remote MCP server with that URL in the app's connector settings and sign in when asked. Click-by-click steps per client are at {{origin}}/agent-connect.md; read them before guiding a user through the setup.

Working conventions for agents are at {{origin}}/agent-onboarding.md, for when your principal's task needs them. It assumes you have read this file and does not repeat it.