Jat

Jat's core bet is architectural rather than model-driven: instead of making one agent smarter, it makes many agents safe to run at once. The mechanism is concrete. Work is partitioned by file path, so an auth refactor agent on src/auth/** and a billing agent on src/billing/** never queue behind each other or clobber each other's edits. Leases enforce that at runtime, signals carry coordination between running agents, and jt (jat-task) supplies a dependency graph so downstream tasks wait on upstream ones. jat-send gives the fleet a shared mailbox so agents can drain an inbox and message each other mid-run — this is the piece that turns parallel workers into something closer to a team. The live board and per-action audit log mean you can actually supervise it; when a run goes wrong you have a record of which agent did what, which is exactly what you lack in a homegrown loop. The family design is the other half of the story. Jat is not one binary pretending to be a platform — it is seven tools that each stand alone and assume each other when combined. jst scaffolds the SvelteKit app your agents work in. jat-secret keeps credentials in an encrypted vault synced across the fleet rather than in environment files on each box. jat-pilot automates your own browser under consent-scoped control, which is the honest way to describe browser automation with a human in the loop. jat-appliance turns a machine into a private compute and inference node, and jat.run pools appliances behind one tailnet endpoint, so the heavy compute question gets an answer that does not involve an API key and a bill. jatstation puts an opinionated agentic Linux desktop at the workstation layer. Where it fits: large codebases, parallel workstreams that split cleanly along directory boundaries, teams with an audit requirement that keeps the code and the logs on their own network, and infrastructure people who already run a tailnet and are comfortable with curl-piped installers. Where it does not: anyone who wants an in-editor AI assistant like Cursor, anyone without terminal and self-hosting comfort, and teams that need a managed service with zero ops. The alpha label is doing real work here — the docs and install path exist, but you should expect to operate this rather than consume it. The scrape names no underlying models for the agents, so model quality and reliability are not verifiable from the published material and should be treated as your problem to evaluate. Likewise, the public material describes a single-command install but not a pricing model, and our run did not reach a pricing page, so nothing about tiers or plans is asserted here.