Feature request: manage ChatGPT Project Sources directly from CLI and MCP

[vgorlovi]

Feature request: manage ChatGPT Project Sources directly from CLI and MCP

Problem

Oracle can already target a ChatGPT Project URL and upload files into a chat, but that is not the same as managing the project's persistent Sources tab.

Today, chat attachments and project Sources are different flows in the ChatGPT UI:

• --chatgpt-url selects the project/folder target
• --file provides local files
• --browser-attachments controls chat delivery
• none of that guarantees that files are added to the project's persistent Sources list

That gap makes it hard to use Oracle as a reliable project-context manager for ChatGPT Projects.

Requested functionality

Add first-class Project Sources management in browser mode for both:

• CLI
• MCP

Suggested operations:

• add
• delete
• replace
• sync

Behavioral requirements

• Target the project's Sources tab directly, not a chat conversation workaround.
• Bypass the OS file picker the same way Oracle already bypasses it for chat file uploads.
• Prefer the Sources panel's own hidden file input / upload surface, not the chat composer input.
• Support headless / terminal-driven execution.
• Batch uploads at 10 files per batch to match the current ChatGPT Sources UI limit.
• Verify success from the Sources list itself:
  • added files appear in Sources
  • deleted files are absent
  • replace/sync converge reliably

Why this matters

A ChatGPT Project is much more useful when Oracle can manage its persistent context directly instead of only attaching files to one conversation.

Offer

I have a working implementation for this locally, including:

• shared browser automation used by both CLI and MCP
• direct Sources-tab targeting instead of chat uploads
• batch handling for the 10-file upload limit
• add/delete/replace/sync behavior
• updated docs/tests

If you're open to it, I can submit a PR based on that implementation.

[goldengrape]

That would be a great feature. However, there is a limit on the total number of files a Project can hold.

My current workaround is to compress a group of related files into a single ZIP archive (for instance, a set of "skills" that need to be reused across the project). I then specify the ZIP filename in the prompt and instruct ChatGPT to decompress and use the contents.

[pdurlej]

Opened a focused draft PR for a safe v1 here: #173. It adds non-destructive project-sources list|add plus MCP project_sources with dry-run and explicit mutation confirmation.

I kept delete/replace/sync out of the first PR on purpose so we can harden the browser DOM path before adding destructive operations. Credit to @vgorlovi and #132 for identifying and proving the Project Sources workflow.

[steipete]

Closing this now that #173 landed as 3e7a5447.

The merged v1 covers the core request with a deliberately safe surface:

• CLI: oracle project-sources list|add --chatgpt-url <project-url>
• MCP: project_sources with operation: "list"|"add"
• add supports dry-run planning
• MCP add requires confirmMutation: true unless it is a dry-run
• browser automation targets the ChatGPT Project Sources tab directly and does not start a consult or select a model

Live verification covered dry-run add, live list, and live add against the steipete ChatGPT project. During live testing, we also hardened the Sources scraper so ChatGPT composer/model/voice controls are not mistaken for source files.

I am leaving destructive delete/replace/sync out of this issue intentionally. Those are still plausible follow-up features, but they should be tracked separately after more live UI coverage because Project Sources mutations are persistent and the ChatGPT DOM is brittle.