Using it
ChatGPT and Claude Desktop
Reach your project's memory from a chat app, for the work that isn't in an editor.
Not all the work is in an editor. Writing the postmortem, planning next quarter, drafting the announcement — these go better in a chat app that can still see what actually happened.
Claude Desktop
Configure it like any other client. The connection file is
claude_desktop_config.json:
- macOS —
~/Library/Application Support/Claude/ - Windows —
%APPDATA%\Claude\
The exact contents are in Set it up by hand; pick Claude Desktop in the picker.
Claude Desktop has no managed instruction file, so its ContextStream capabilities and retrieved project context arrive through the connection without the editor-specific initiative rules. Ask directly: "check what we decided about the billing migration".
ChatGPT
Through the ContextStream Custom GPT in the GPT Store. You connect your account once, then it can read your workspace's memory in any conversation.
What these are good for
- Writing about the work. A postmortem grounded in the actual sessions beats one written from memory.
- Planning. "What's still open on billing, and what did we learn from the last migration?"
- Catching up. Monday morning, before opening an editor.
- Explaining to someone else. Drafting the message that explains a decision, with the decision in front of it.
What they're not for
Editing code. Neither has your filesystem. Use a coding editor connected through the ContextStream MCP for changes.
If it doesn't work
Claude Desktop doesn't show the connection. Quit and reopen — it reads config at startup only. The config path is easy to get wrong; check it against the picker.
The Custom GPT can't see my workspace. The account link expired. Reconnect from the GPT's configuration.
It answers generically. Ask explicitly for project context — "check our decisions first". With no instruction file, the prompt has to carry that.