anthropics / anthropics/claude-code

Claude Code has no way to read/import claude.ai Project (Cowork) knowledge files

オープン
#91,425 コメント 2 件 リアクション 1 件 担当者 0 名 GitHub で見る
area:cowork enhancement
主要言語
Python
スター
145k
フォーク
23.1k
PR マージ指標
PR 指標を取得中

説明

## Description

Claude Code (CLI) has no tool or connector that can read the contents of a claude.ai Project — specifically the "context files" attached to a Project (a Cowork project, in this case), which function as a shared knowledge base for that project.

**This significantly undercuts the value of "sharing" a Project.** Sharing a Cowork/Project with a colleague is presumably meant to let them work with that context, not just view it once in the claude.ai UI. But today: the colleague can open the shared Project in a browser and read every file — the share works fine at that layer — yet has no way to bring that same content into *their own* Claude Code session. They're stuck manually opening each file and copy/pasting its content one at a time. For a project with more than a couple of files, that's not really "sharing" in any useful sense for Claude Code users — it's "here's something you can look at, by hand, but not use."

Confirmed no existing path works:
- `WebFetch` only authenticates against `claude.ai/code/artifact/{uuid}` (published Artifacts) via the user's claude.ai login. A `claude.ai/project/{uuid}` URL returns only the unauthenticated SPA shell (page title "Claude", no content) — even though the requesting user has legitimate, shared access to that Project and can see its full contents in a browser.
- No MCP connector among the available first-party connectors (Atlassian, Microsoft 365, Slack, Notion, etc.) exposes claude.ai's own Project/file storage.
- `DesignSync` is a different product surface (Claude Design component-library projects), not general Projects with markdown/HTML knowledge files.

**Workaround: Claude in Chrome.** Driving the Project page via Claude-in-Chrome browser automation can extract file contents, but this is a poor substitute — it's a UI-automation path (slow, brittle, extension-gated), not a reliable substitute for a direct read tool, and isn't available in every Claude Code environment (e.g. headless/remote/CI-style sessions).

## Repro steps

1. User A creates a claude.ai Project (or Cowork project) with several "context files" (`.md`/`.html`) attached, and shares it with User B.
2. User B opens the Project in a browser — confirms they have real, working access: all files readable, normal Project UI.
3. User B, in a Claude Code session, asks Claude to fetch/read/import the contents of that same Project via its URL (`https://claude.ai/project/{uuid}`).
4. Observe: no tool succeeds for User B either, despite their legitimate shared access. `WebFetch` returns an unauthenticated shell; no MCP connector applies. The only workarounds are (a) opening each file in the browser and pasting its content into the Claude Code session by hand, one file at a time, or (b) Claude-in-Chrome browser automation, which is unreliable and not always available.

## Expected behavior

Some supported way for Claude Code — running as a user who owns *or was shared* a Project — to read that Project's attached context files, analogous to how `WebFetch` already authenticates to the user's own published Artifacts. Sharing access to a Project should mean shared *usable* access, not just shared *viewable* access.

## Impact

Blocks a reasonable, increasingly common workflow: a PM (or any Cowork user) maintains working documents as Project context files and shares the Project with an engineering collaborator so they can work from the same material in Claude Code. Today that collaborator gets nothing but view access — they still have to hand-copy every file into their own session, or fall back on a fragile browser-automation workaround. Encountered with a real project containing 16 context files, where copy/pasting each one individually is a genuine workflow blocker, not a minor inconvenience.

## Environment

- Claude Code CLI
- Reported from a session using model `claude-sonnet-5`

コントリビューションガイド

このリポジトリのコントリビューションガイドは索引されていません

調査の方向性

The issue names Claude Code CLI, WebFetch, MCP connectors, and claude.ai Project URLs but no source files or tests. Start by tracing the CLI's WebFetch and MCP integration points and reproducing access to a shared Project URL. Done means an owned or shared Project's attached context files can be read in Claude Code without manual copying or browser automation.

索引モデルが issue の本文から書いたものです。

評価

技術スタック
python
領域
api, cli
issue の種類
機能追加
難易度
5/5
見積もり時間
1週間以上
活発さ
活発
明瞭さ
説明が足りない
初心者へのやさしさ
35/100

新しい issue をメールで受け取る

初心者向けの GitHub issue を短くまとめたダイジェスト。