FAQ

Does this work without Tailscale?

Yes. Tailscale is just the easiest way to give the laptop a stable, private route to the remote. LAN works the same way. So does any VPN. The CLI doesn’t know about Tailscale beyond setup’s peer-discovery convenience.

Does this work outside macOS?

The CLI runs on any platform with Node ≥ 20, git, rsync, and ssh (so: macOS and Linux).

The agent server runs anywhere too. The install subcommand for launchd is macOS-only, on Linux, write a systemd unit (one is shown in Running the agent).

Can the agent run on Linux instead of a Mac?

Yes. We just need claude, git, and Node. The “remote” framing in docs is the common case, not a hard requirement.

What happens if the laptop and remote are out of sync (different commits)?

The agent doesn’t care about your laptop’s git state. It only cares that its own working tree is clean before each run. The diff it returns is computed against whatever’s on the remote. As long as your laptop tree is at the same commit as the remote (which is what rsync --delete enforces), git apply will work.

If you’ve committed locally between sync and apply, the patch may still apply, git apply is line-based, not commit-based. If it doesn’t, you’ll see a git apply --check failure and the patch is saved to .patchwire/last.patch.

In v1, no isolation between laptops/users. Two callers using different project names won’t collide on disk, but they share PW_AGENT_TOKEN and there’s no per-call concurrency control inside one project. Multi-user is on the roadmap.

Is the prompt sent to Anthropic?

Yes, by claude running on the remote, exactly as if you ran claude locally. Patchwire doesn’t add any third party to the picture. See Anthropic’s data policy.

Can I use a different LLM CLI?

Yes, in principle. The agent just spawns PW_AI_BIN with PW_AI_ARGS and pipes the prompt to stdin. Any tool that edits files in place and exits 0 on success will work. We’ve only tested with claude.

Does this stream output?

Not yet. v1 is request/response, you wait for the whole run, then see the diff. Streaming is on the roadmap, but harder than it sounds because we don’t know what changed until git diff runs at the end.

How big can the project be?

Tested up to ~1k files post-excludes. The bottlenecks are (a) Claude’s own context handling and (b) initial rsync time. Incremental syncs are tiny regardless of repo size.

Why one-way sync?

Bidirectional sync sounds nicer in theory and is a nightmare in practice. Conflicts, “.sync-conflict-X.dart” files, surprise deletes. The diff-back model is strictly better for this use case: you always know exactly what’s changing, and you decide when.

How do I uninstall everything?

# laptop
pnpm remove -g @rebink/patchwire
rm -rf ~/.patchwire

# remote
patchwire-agent uninstall
pnpm remove -g @rebink/patchwire
rm -rf ~/.patchwire

Per-project: delete patchwire.yml and .patchwire/.

Can I use this from a CI runner instead of a laptop?

Technically yes, the CLI is just node + commander. But CI is a weird fit: the diff preview prompts are interactive, and the value prop (local IDE stays fast) doesn’t apply. If you have a use case for non-interactive AI runs in CI, open an issue and we’ll talk.

Does `--save-only` skip the apply prompt?

Yes. The patch goes to .patchwire/last.patch and the command exits. Apply later with patchwire apply.

How do I see what `claude` actually output?

The CLI prints a summary; the full stdout is in the JSON response. With PW_VERBOSE=1, you’ll see more. Or hit the API directly (see HTTP API).

FAQ