Cersei

The deciding trait is the return value. You call a builder, hand it a provider and a tool set, and you get the result of a run back inside your own program, instead of spawning a process and scraping its output. If you have ever wrapped a coding agent in a subprocess and regretted it, that one difference is the entire pitch.

What you give up is reach: you have to be writing Rust, on macOS or Linux, and nobody hands you a finished product. Pick it if you are building the thing. Pick a terminal agent if you are only trying to finish today's ticket.

The risk is derivation without a contract. The architecture is a port of somebody else's agent, which means the original can change its prompts, its tool surface and its file conventions whenever it likes, and this project learns about it afterwards. No compatibility policy is documented for that case.

The dependency line makes it worse. A build takes whatever the default branch happened to contain that morning. What it does right is the permission policy: it is an argument to the builder rather than a prompt the model can talk its way past, so the caller decides what may be touched before the run starts.

1. The comparison measures engineering properties rather than capability. Startup time, memory footprint and throughput say nothing about whether a task was completed correctly, and no evaluation of correctness is offered anywhere. 2. The harness that produced the figures is authored by the same project, so the numbers are an assertion in the shape of a measurement.

2. The more interesting claim carries no number at all. Memory is offered in two forms, a file on disk and a graph, and the graph is the default in the shipped binary. That is a real decision about how prior work is retrieved, and it is stated rather than demonstrated.

The ownership question answers itself. A single individual holds this, there is no entity, no hosted tier and no revenue line, so the eighteen-month question collapses into a question about one person's calendar. Four hundred stars is evidence of interest and finances precisely nothing.

Pricing power: none, and none sought. Moat: none either, since there is no hosted component, no proprietary data and no distribution channel, so nothing accrues with use. Likely path is that the ideas get absorbed by a larger project while this one slows. Position: use it as a dependency and accept that you are the support contract.

Cost is not the interesting question. Sixty seats at zero, with inference billed to accounts finance already reconciles. Governance is the interesting question, and the answer is that there is none: nothing federates identity, nothing enforces a policy across machines, and nothing writes down what an agent did in a form that survives the session.

In a pipeline it is a flag, not a product. Unattended operation means switching permissions off and parsing structured lines out of standard output. Onboarding is a systems engineer's week rather than a mid-level engineer's afternoon. Not yet, unless it sits buried inside a service my platform team already operates.

The licence is the short version of ownership and this one is the permissive kind, so a fork is mine outright with nobody to ask. Better than that, the provider list treats a box on my desk the same way it treats a hosted API. Point it at Ollama or vLLM and nothing crosses the network. I did not have to earn that with a proxy.

Client-side MCP means servers I already run attach as tools without a shim. What I do not get is the other direction: it speaks the protocol but does not serve it, so nothing else in my toolchain can call the agent back.