Claude Code Haha

This is a desktop shell around a coding agent, and the trait that decides it is git handling. You choose a branch, decide whether the session works in your current tree or an isolated worktree, and then review the turn file by file in a syntax-highlighted diff before anything sticks. Global search across every session is a keystroke away.

The catch is that it is a wrapper, so you inherit its bugs on top of the agent's. Pick it if you run several sessions at once and want them visible. Pick Claude Code directly if you would rather keep the stack thin.

The risk is surface area against maintenance capacity. The author states the project is maintained in spare time, and the feature list runs to computer control, a third-party skill marketplace, phone access, orchestration scripts the model writes at runtime, and animated pets. Every one of those is code that touches your machine, and no single maintainer audits all of it at the pace it is being added.

Install only what you use and leave the marketplace alone. What it does right: every model request is logged locally with status and timing, so a stuck or failed call is diagnosable rather than mysterious.

1. Actions are applied to a working tree the user selected, so the unit of change is a turn rather than a file. 2. Verification happens in two places: the changed files are presented as a reviewable diff with an undo for the whole turn, and the resulting page loads in an embedded browser carrying session state, which closes the loop without a separate deployment. 3. Orchestration is generated rather than fixed, with the model writing scripts that drive subagents and support interrupts and resume.

No benchmark is offered, and for a shell around someone else's agent that is the honest position.

Follow the money and it goes sideways. Development is supported by sponsorships from API relay services, each entry carrying a referral link, a promotional code and a signup credit, alongside payment buttons for direct donations. That is an audience being monetised through affiliate spend rather than a product being sold.

It works while attention lasts, and attention in this category lasts about eighteen months. There is no entity, no contract and no roadmap commitment to buy. Likely path: the author accepts sponsorship for custom work, or the repository slows. Position: use it, do not put it in a procurement document, and keep the underlying agent installed separately.

The demo is genuinely good, and then the integrations list arrives. Sessions can be driven and approved from five different messaging platforms, and a scanned code continues the session in a phone browser while the desktop is locked. That is source code and approvals crossing onto personal devices and consumer chat services, which is a conversation security will finish quickly.

Installers also surface signing warnings that ask a user to click through a system block, which is precisely the habit we spend money teaching people not to have. No central administration, no audit. Not yet.

MIT, and the extension points are not an afterthought. Servers are added in a visual manager that handles stdio, Streamable HTTP and SSE, and each one is scoped to a project, shared, or global, which is the scoping I usually have to fake with symlinked config files. The model list takes the hosted vendors, preset providers, or a local endpoint, so my own box is a first-class choice.

For debugging there is a source path: install with bun, copy the example environment file and run the command-line entry point directly. Grudging respect for a wrapper that ships its own internals.