Starpod

The deciding trait is one agent per directory, entirely self-contained. Everything it knows lives beside the work it knows about, so your notes on one project never leak into another and deleting a folder deletes the assistant with it. That containment is unusual and it makes the mental model pleasantly simple.

It is a personal-assistant runtime rather than a coding tool, so it will not review your diffs or manage your branches. Pick it if you want a resident helper per project. Pick a coding agent if you wanted one that ships code.

Those two features together are the failure mode. Self-modification means the program that runs on Tuesday is not the one you reviewed on Monday, and scheduling means nobody is present at the moment it changes. Nothing in the row describes approval before a skill is written, a record of what changed, or a way to pin the set.

What it does right is keep the blast radius on disk. Everything a given agent holds sits in one directory, so inspecting or discarding the whole state is a single operation.

1. Keeping the authoritative record in plain text and the index derived from it is the correct ordering, because the index can be rebuilt and a corrupted embedding store cannot be read. A person can audit what the agent believes. 2. Full-text search rather than vectors is a deliberate choice with a known precision profile, which is easier to reason about than a similarity threshold nobody published.

2. Partitioning memory per user makes isolation a property of the schema rather than of prompt discipline. 4. No evaluation is published.

The cloud option is the business plan showing through. A local runtime given away with a hosted deployment beside it is the standard structure, and it works when operating the thing is annoying enough that people pay to stop. Whether that holds for a personal agent is untested, and seventy-eight stars is not yet a test.

Moat: none today; the hosted operational knowledge is the only candidate. Likely path: the company pivots toward whichever of its products finds users first. Position: use the local runtime, and do not plan around the hosted one until it has a price.

The vault is worth naming. Secrets in an encrypted store rather than plain environment files removes the most common way a developer machine leaks a provider key, and it is the kind of detail that shortens a security review rather than extending it. Somebody here has answered a questionnaire before.

The rest is the usual gap: no identity integration, no provisioning, no central audit across sixty installations, and nothing that runs inside our delivery process. Not yet, though I would revisit it if a managed deployment ever arrives with a console attached.

A single init command producing a self-contained agent directory is the right ergonomics, and installing from the language's own registry means no curl into a shell if I do not want one. The licence is permissive, the source is Rust, and the state is files I can read, move and version.

The gap is that the only part not under my roof is the part that thinks. No local endpoint, so every scheduled run leaves the machine, and no MCP client, so the servers I already run are not available to an agent that was otherwise designed to live here.