Is it safe to run on my real repo? audit only reads. apply is a dry run unless you pass --yes, changes only what is missing, never deletes or overwrites,
and saves an undo snapshot, so restore can put it back. It has been exercised against the real API on a scratch repo (see the README), not on every kind of repo.
Why doesn’t it protect main with pull requests? Because with one maintainer that blocks your own merges. See why-solo.md. It blocks deletion and force-push instead.
What if I work with one other person? Then required reviews can make sense. solo-blocker only fires when you are the only account with push access; with two or more it reports ✅ “team repo”.
github-solo simply never adds such a rule.
Does it send my data anywhere? No. It talks only to the GitHub API (or to SOLO_API_BASE if you set it, for testing). The token is never printed or stored.
Which token do I need? A classic token with repo, or a fine-grained token with Administration (read/write), Contents, Metadata and Pages. See the README. solo.py doctor OWNER/REPO shows what your token can read.
Why does a row say “could not read (HTTP 403 …)”? The token or the plan cannot read that setting, or a sandbox/proxy in front of the API refuses the path. The row is a ⚠️ instead of a guess. references/troubleshooting.md lists the causes.
Why are some items ➖? Private repos on a free plan have no Secret Protection, Code Security or rulesets. They are reported with the reason and never enabled.
apply says “not a clone” and skips my files. Generated files are written locally only inside a clone of that repo. Run it from the clone, or add --commit-files to commit through the API
(it refuses when a push to the default branch may trigger a release workflow, unless you add --accept-release-risk).
Does it work for organization repos? It uses the same API calls, but it has not been tried on an organization-owned repo. Organization policies can change answers (for example a 409 on the workflow permissions endpoint).
Windows? Yes: the tests run in CI on Windows and the live checks were run from Windows. See recipes.md for the usual snags.
Which agents can use it? Anything that reads SKILL.md or AGENTS.md, or speaks MCP. Verified: Claude Code (the CLI driven from Claude Code sessions on Linux and Windows, plugin manifest validation, MCP connection). Codex and Cursor setups follow their documentation but were not run.
Can I add my own check? Yes: a check is one function in scripts/solo.py returning R(...) rows, plus a planner in plan_for if it can be applied. See CONTRIBUTING.md.