dev-read-gh-public-repo

Read files from a PUBLIC GitHub repo that the Claude Code web/remote environment reports as out-of-scope for its GitHub MCP tools. Use when (1) running in the web/remote container (claude.ai/code, NOT

SKILL.md

GitHub

Read a public GitHub repo that's out of MCP scope

The problem

In the Claude Code web/remote environment, the GitHub MCP tools (mcp__github__*) are locked to an allow-list of repos chosen when the session was created. Reading any other repo — even a public one — fails like this:

Access denied: repository "owner/name" is not configured for this session.
Allowed repositories: <only-the-scoped-ones>

The fix — try in this order

1. Add the repo to session scope (primary path)

Call the platform's add_repo tool (mcp__Claude_Code_Remote__add_repo in this container) with the owner/repo — this is the supported way to widen scope mid-session (see github-ops.md, "Repo scope": "For cross-repo work, add the target repo via the platform's add_repo"). On success it returns a clone command; run it, then call register_repo_root so the harness picks up the new repo. After that, the normal mcp__github__* tools / Read / Grep work against it like any in-scope repo — full read access (issues, PRs, file contents), not just a one-off file dump.

If add_repo is unavailable in this session, or the call is denied (authorization/policy error, GitHub App not installed for the repo), fall through to step 2.

2. Codeload tarball fetch (fallback)

Fetch a tarball over plain HTTPS instead of using git — the same bootstrap pattern dev-setup-webenv uses to pull claude-resources into a web session:

cd /tmp && curl -fsSL "https://github.com/<owner>/<repo>/archive/refs/heads/<branch>.tar.gz" -o repo.tar.gz \
  && mkdir -p <repo> && tar -xzf repo.tar.gz -C <repo> --strip-components=1
# now Read / Grep files under /tmp/<repo>

Try the actual default branch first (commonly main, then master if that 404s).

3. Plain git clone — verify before relying on it

cd /tmp && git clone --depth 1 https://github.com/<owner>/<repo>.git

Whether this works depends on how this session's git proxy is scoped, and that has been observed to vary: some container configurations reject git clone/git ls-remote of an out-of-scope repo with a 403 (the assumption documented in dev-setup-webenv), while a plain git clone of a public, out-of-scope repo has also been observed to succeed outright in-container. Don't assume either outcome — check first with git ls-remote --exit-code https://github.com/<owner>/<repo> HEAD (expect a commit hash, not an error) before depending on it. Steps 1–2 above don't depend on this uncertain behavior, so prefer them.

Gotchas

  • HTTPS only — SSH fails. There is no ssh binary in the container, so

    git clone git@github.com:owner/repo.git dies with ssh: not found. Always use the https://github.com/... URL (or the codeload tarball URL).

  • No auth needed for public repos over HTTPS — the fetch just works, assuming

    the environment's outbound network policy allows github.com / codeload.github.com.

  • Clone/extract outside the working repo (/tmp, or any path that isn't the

    user's project) so the checkout stays ephemeral and never gets committed by accident.

  • --depth 1 (or the tarball, which has no history at all) keeps it fast; only

    fetch full history if you actually need it.

Limits — don't misapply

  • Reading only. This is for research: reading a doc/spec, copying a pattern. It

    grants no write access and is not a substitute for proper repo scope.

  • Public repos only. Private repos still require adding the repo to the session

    scope via add_repo (or real auth) — a plain HTTPS clone/tarball fetch will 404 or prompt for credentials and fail.

  • Web/remote env only. On a normal local session, just use gh or an existing

    checkout instead — don't reach for this.