Lemma

The deciding trait is that it drives the coding agent you are already paying for. Pair a machine, and Claude Code or Codex or Cursor, already logged in on that box, does the work when a job fires. You are not buying a second inference bill or learning a second agent's habits, which is the part that usually kills these platforms in week two.

The catch is that you are now operating a platform, not using a tool. Pick it if you want unattended work driven by the agent you already trust. Pick a plain cron job and a script if you only have one thing to run.

The execution substrate is a laptop. Agent Host pairs a machine and dispatches work to it, which means a run succeeds or fails on whether that machine is awake, online and still authenticated. A platform that keeps working between sessions is being asked to survive somebody closing a lid.

Nothing in the row documents what happens to a dispatched run when the pairing drops mid-task, and there is no sandbox around the shell it is driving. What it gets right is honesty about the boundary: server-run agents and paired-machine agents are separate things, and the docs do not pretend otherwise.

1. The system is authored by the coding agent and then verified by the CLI, which is an unusual and defensible division: generation is probabilistic, import is not. Apps, tables, workflows and permissions land as files, so the artefact under review is text a human can read rather than state inside a running service. 2. That makes the whole configuration diffable, which is the property most agent platforms give up first.

2. Verification is described as an import step, and no evaluation of the generated systems is published. The design is sound; its accuracy is untested in public.

The business model is legible, which is rarer here than it should be: self-host free, pay for Lemma Cloud when you would rather not run it. That is the correct shape for infrastructure, and it is also the shape with the thinnest margins, because the escape hatch is always a server you already own.

Moat: switching cost, eventually. Once a team's automations live here, moving them is a migration project rather than a decision. Likely acquirer: a platform vendor that wants the orchestration layer above the coding agents it does not own. Position: workable, and 387 stars means the thesis is unproven.

Sixty seats cost nothing in licence and everything in operations: we host the server, we patch it, we own the uptime, and nobody is on call for it but us. Budget it as a service my platform team runs, not as a tool I buy.

The trigger surface is the part I like. Work starts from a schedule, a webhook or a table event, and approval steps can sit in the middle, which is the first thing procurement asks about automation touching production. No directory integration or audit export is documented. Approved with conditions: server-run agents only, and human approval on anything that writes.

AGPL-3.0, which is the licence that actually means something: anyone offering this as a service has to ship the source back. I like that more than I like MIT here, because the thing being protected is a platform, and platforms are what get quietly reskinned into someone's SaaS.

Model routing takes any Anthropic- or OpenAI-compatible endpoint, so a gateway on my own hardware is a URL change rather than a fork. ACP is a published protocol rather than a private wire format, which means the pairing is inspectable. No MCP client, so my existing servers stay outside; that is the gap I would close first.