How Unl works
Unl, the open-source why agent for your coding agents. This film walks through it in real browser recordings, with an AI narrator. Every chapter and every word it says is written underneath.
npx unlimitless initFree. Apache-2.0. No Unl account. Requires Node.js 20 or later.
Chapters
- 0:00 A decision with a reason
- 0:36 Before the agent acts
- 1:18 What the record carries
- 1:59 From words to visible work
- 2:39 Local use and its limits
- 3:22 Start with your own project
Transcript
0:00 A decision with a reason
On screen, the public Loops page, a recorded October example.
This is a real decision in Unl's public record. Neil said an agent must not throw away work it did not make. The reason matters: unsaved changes might be the only copy. The page connects that decision to a build, a report, and a later verification. We are looking back at a recorded October example, not staging a new incident. Unl is the open-source why agent for your coding agents. It keeps what you decided, together with why, so that your reasoning can reach an agent before it acts. The human still decides what holds.
0:36 Before the agent acts
On screen, the public view of the hosted fleet.
Here is the public window onto the hosted fleet. A Codex turn appears beside the decisions served to it. Opening a reference lets us read the instruction and its reason. That gives the agent something more useful than a detached rule: it can understand what the person was protecting. The agent still has to reason about the task. A served decision is not proof that every action followed it, and a report is not the same thing as independent verification. Those distinctions matter. The record makes the connection available to inspect; it does not make the model infallible. This hosted fleet is one working example, not a promise about every installation.
1:18 What the record carries
On screen, the decision format documentation in the open repository.
The public format documentation describes the pieces in plain language: the decision, the reason, alternatives ruled out, where it applies, who chose it and when, and what it replaces. The JSON shown here is a documentation example, not a customer's decision. Keeping the reason makes it possible to revisit a choice without guessing what the earlier person meant. Keeping replacements means an old answer need not masquerade as the current one. The format is open. The important boundary is human authority: an agent can propose a choice, but that does not make it canonical. The person's own confirmation does.
1:59 From words to visible work
On screen, the public Loops page, a second recorded October example.
This second historical example concerns the connection page. Neil wanted every supported tool offered in the same place, because a tool arriving later should not become harder to find. The public record links his words to the build and the agent's report, and separately shows a verification. That is the practical loop: say what you want, keep why it matters, let an agent work, then inspect what happened. The reason also helps with the next change. If another tool arrives, the agent can use the same intent instead of making the person explain the layout choice again. The record is useful because the work continues.
2:39 Local use and its limits
On screen, the open repository's README, how it works and its limits.
The local product keeps a project record on your machine, with readable decision files and sections your coding tools can use. You do not need an Unl account to begin. Local does not mean no information can leave the machine: if your setup asks an AI provider to select relevant decisions, that provider receives context. The documentation also states limits. Relevance can be too broad, and integrations do not all have the same verified coverage. A before-action check is not a security sandbox or a replacement for tool permissions. We are showing the mechanism and its documented boundaries, not claiming a benchmark victory or that the system always chooses perfectly.
3:22 Start with your own project
On screen, the connection page.
The connection page separates the free local start from the hosted connection choices. They are different paths; they are not automatically one shared record. For a local project with Node twenty or later, the starting command is npx unlimitless init. Follow the setup for the coding tool you actually use. Begin with a decision you have already made and the reason you made it. Check what reaches the agent, and check the resulting work. You can change your mind; the record should carry that change forward. Unl keeps the why available. Your agent does the work. You remain the person who decides.
The recordings are real browser captures of public pages, not staged sessions. The two Loops shown are recorded October examples, looked back at rather than replayed. The narrator is an AI voice describing the product; it does not speak as any person.
npx unlimitless initFree. Apache-2.0. No Unl account. Requires Node.js 20 or later.