Crystal

You should not adopt this today, and the reason has nothing to do with how it feels to use. Maintenance stopped: the publisher deprecated it in February 2026 and pointed users at a replacement, so the trait that decides it in daily use is the one you notice six months in, when a broken integration stays broken. Everything below that line was competent and remains competent.

Pick Conductor if you want parallel sessions with an active maintainer, or move to the successor the vendor named. Install this only if you specifically want the version you already know and can live without fixes.

Two problems compound. History-rewriting operations are exposed as ordinary application actions, and history rewriting is the class of git mistake that is expensive to reverse. On top of that sits an abandoned desktop runtime, so anything found in its dependencies from now on stays found. An unmaintained application that can rewrite your branches is a risk that grows while you do nothing.

What it did right: each session carried its own run script, so verifying an attempt meant executing the project's own commands rather than trusting what the agent said it had done.

The design deserves recording even though the product stopped. 1. Competing attempts at the same task ran concurrently and in isolation, which converts model non-determinism from a complaint into a sampling strategy. 2. Comparison happened on tracked diffs, so selection used the artefact rather than the narrative around it. 3. Sessions persisted, so an attempt could be revisited instead of re-run.

No evaluation was ever published, which is consistent: the tool measured nothing and let its user do the measuring. That division of labour is the part worth copying.

The company did not fail; the product was retired and a differently branded successor took its place with the same isolation model underneath. That is a repositioning, and repositioning usually means the original could not be charged for. Three thousand stars of goodwill were spent to acquire the audience for the next attempt, which is a legitimate use of open source and a poor basis for a dependency.

There is no acquirer for a deprecated desktop wrapper. The interesting question is whether the successor carries a price. Position: exit, and watch the replacement's pricing page for what the vendor learned.

Sixty seats at zero is not a bargain when the vendor has stopped. Packaging tells the same story: a signed macOS build and a cask, while Windows meant compiling from source and Linux was never documented, so a standard image was never on offer. There is no headless mode, so it contributed nothing to a pipeline, and no directory integration or audit trail to describe to security.

Onboarding cost is now irrelevant. Not yet, and for this product that means not ever. Anything already installed on a company machine should come off the next time we touch it.

This is the case permissive licensing exists for. The vendor walked away and the source did not, so I can clone it, run the setup script and build the desktop app myself with no permission from anyone. It drives agent CLIs I already authenticate separately, which means nothing here holds credentials and a fork inherits no secrets problem.

What I do not get is extensibility: no MCP in either direction and no model choice of my own, because the agents underneath make those decisions. I would fork it for the worktree handling and throw the rest away.