Glue · part of Supernova
Glue is the part of Supernova that reads code. It reads a codebase’s history and its source, and tells you in plain facts who built what, where the knowledge sits and what depends on what. You can use it in two ways: to understand any codebase, or as the first step in building the next phase of a product that already exists.
Two ways to use Glue
Part 1 · For any codebase
For a team taking over a product, a company checking what it is buying, or a new technical lead. Antino invites you as a guest. You connect a GitHub repository, and Glue shows you what its history says about who built it and how it changed.
Part 2 · Inside Supernova
When Antino takes over a product another team built, Glue is the first step. It reads the full history and every TypeScript and JavaScript file. Supernova’s agents write the specification back out of the code. We add the new phase alongside it. Supernova plans the work, and Antino’s developers build it.
Part 1 · Read any codebase
You do not need to install anything or give Glue access to your servers. Glue reads what GitHub already holds about the repository, and nothing more unless you ask us to read further.
You can connect up to three repositories, and we can raise that. You can remove a repository yourself at any time, and everything Glue read from it, and your token, is deleted with it.
Ask for access on this page. When someone at Antino confirms it, you get a sign-in that opens Glue and nothing else in Supernova. It has an end date. When the date passes, access ends everywhere at once, even in a window you left open. We can extend it.
Paste a GitHub address. Glue asks GitHub first, before storing anything. If GitHub will not show it the repository, Glue tells you in plain words what to check. Glue also needs a GitHub token, a key GitHub gives you so an app can read on your behalf. For a public repository, make one with nothing ticked: it allows nothing, and only stops GitHub from limiting how often Glue can ask. For a private one, make a read-only token for that one repository. Tokens are encrypted before they are stored and never shown again.
In under a minute, Glue reads up to the 3,000 most recent commits on the repository’s default branch, leaving out merge commits. You see who changed the code, how often and when. If the history is longer than that, Glue says so rather than pretending it read everything.
Glue shows who made the commits, how many each, and how the work changed over time. It reads the history, not your files: Supernova does not keep your code. The full eight-section report shown below has so far been produced for Express.js.
Part 2 · Build the next phase
Many products Antino takes on already exist. The documents are out of date or missing, and the code is the only accurate record of what the system does. So we start from the code and work forward, in four stages.
Glue reads the codebase's full history and its TypeScript and JavaScript source. No AI is used at this stage: every number comes from Git, GitHub and the compiler.
Glue · no AI
Code agents read each feature out of the source and write it up as use cases and user stories. Each one is stored with the functions it was read from, so it can be checked against the code.
Reviewers read the specification and write notes: what is wrong, what changed, what everyone knows and nobody recorded. Each note is traced across the whole specification, and the changes it calls for are saved as new versions that can be rolled back.
The new phase is written alongside what already exists. Supernova then plans it the way it plans any project, without replacing what is there: use cases, user stories and developer tasks.
Stages 02 to 04 are run by Antino’s team as part of a project. The Express.js example below has had stage 01 only, so it shows no specification and no plan.
From stage 02, the code the agents need is sent to the AI provider set up for the project. Ask us which one.
Worked example · a public repository
Express.js is one of the most widely used web frameworks for Node.js, and its code is public, so we can show you exactly what Glue reads. This is expressjs/express at revision 9a34acf, read on 25 Sep 2026. Nothing on it is made up. Antino read the whole history of Express.js from a full copy of its code; when you connect a repository yourself, Glue reads up to the 3,000 most recent commits and tells you if there are more.
Who built this, and who is still here?
Two people wrote 82.4% of the 5,685 commits. Neither has committed since 2 November 2023.
How it is counted. Commits per author email, over the whole history, merges left out.
Check it on GitHub: Commit history at this revision ↗How many people would have to leave before nobody knows a file?
113 of the 141 code files are more than half one person's work. No file needs more than two people to account for half its changed lines.
How it is counted. Changed lines per person in each file's full history, renames followed.
Check it on GitHub: lib/response.js, the most-changed file ↗Where has change piled up, and is it still moving?
Six files, the whole of lib/, have taken 37.0% of every recorded change to the code Glue read (1,119 of 3,020), from 4.2% of its files.
How it is counted. Every commit that touched each file, across all of history.
Check it on GitHub: The lib/ folder ↗What kind of work has the history recorded?
In the last 12 months, 56.1% of commits were tooling and maintenance, up from 32.3% the year before, and an automated dependency bot accounts for the rise. Fixes fell from 23.5% to 7.8%.
How it is counted. The type of work each commit message states, compared year on year.
Check it on GitHub: Commit messages at this revision ↗What is actually in the code, and what is hard to read?
109 functions make up the code Express.js gives its users. All of them sit in 6 files under lib/, and none is longer than 100 lines. The longest, res.send, is 97 lines.
How it is counted. Every function the TypeScript compiler finds, measured in lines. The rest of the functions Glue found are in the project's tests and examples.
Check it on GitHub: res.send in lib/response.js ↗What depends on the code you want to change?
Glue resolves 369 links from one function to another. None of them reaches res.send, which is called from 240 places in 54 files.
How it is counted. Calls the compiler can follow from the caller to the exact function it reaches. Express.js handlers call res.send() on the response object they are given, which a reader that does not run the code cannot pin to one function, so none of those calls is followed. The report says so.
Check it on GitHub: One of the files that calls res.send ↗As a guest you can open the full eight-section report on Express.js inside Glue, with every file, every commit and every function behind these numbers.
The report’s eight sections
Every section answers one question. There are no scores out of ten: every finding comes with the files, commits, people or functions behind it, so you can look at them and tell us if we are wrong.
Answers
Who wrote the code, in which parts, and who is still working on it.
How it is counted
Commits per person over the whole history, and the last 12 months against the 12 before. It also finds files where fixes keep coming back (three or more fixes, on two or more days) and changes that arrived with tests.
Cannot tell you
How hard anyone worked. A commit is a record of a change, not a measure of effort.
Answers
How much each file depends on a few people.
How it is counted
The lines each person changed in a file over its full history, following the file when it was renamed. The bus factor is the smallest number of people whose changes add up to half of the file's changed lines.
Cannot tell you
Who understands the file today. It shows who changed it in the past.
Answers
Where change has piled up, and whether it is still happening.
How it is counted
How many commits touched each file, and its churn: lines added plus lines removed. It also shows which files changed in the 90 days before the version Glue read.
Cannot tell you
Why a file changes often. Rewrites and clean-ups count as churn too.
Answers
What kind of work the history records.
How it is counted
The kind of work each commit message states, such as a fix, a feature, tooling or documentation, using fixed written rules rather than AI. A message that states no kind of work is counted as unclassified, not guessed.
Cannot tell you
Productivity. It reads only the first line of each message.
Answers
What is in the code, and which parts are hard to read.
How it is counted
Every function the TypeScript compiler finds and its length in lines, with a closer look at long or deeply nested functions in the shipped code.
Cannot tell you
Whether code is wrong. Length is a cost to read, not a defect.
Answers
What depends on the code you want to change.
How it is counted
Calls the compiler can follow from the caller to the exact function they reach.
Cannot tell you
Every dependency. A call made through an object or a callback can be invisible to a reader that does not run the code, so a function with no known caller may still be used.
Answers
What the code does, written up so it can be checked.
How it is counted
Produced in stage 02, when the specification is written from the code.
Cannot tell you
Nothing yet for Express.js: that stage has not been run on it, so the section shows only what stage 01 read.
Answers
What it would take to build the next phase.
How it is counted
Produced in stage 04, when the next phase is planned.
Cannot tell you
Nothing yet for Express.js, for the same reason.
How the numbers are checked
Every figure comes from Git, GitHub and the TypeScript compiler, counted by code. AI agents only come in at stage 02, when the code is written up as a specification.
The report names the revision it read, and its links open that same version on GitHub. What you check is exactly what Glue read.
For the Express.js example, a verification suite counts every figure in the report again from scratch, from the stored data and from Git itself, and fails if a single number differs. On 25 Sep 2026 it ran more than 1,000 checks, and every one passed.
Each report says what it does not measure, in the same place as what it does. A number without its limit reads as more than it is.
What the checks prove, and what they do not: they prove each report counts exactly what it says it counts. They cannot prove a definition is the right one. In Express.js, about two in every five commit messages do not say what kind of work they were, and the report shows them as unclassified rather than guessing.
The agents behind it
Supernova is built around 43 named AI agents in service, each with one job. The full roster also lists agents still in beta. These are the ones that take a codebase from Glue's reading to a plan for its next phase.
Each of these agents has a narrow set of tools. The code agents cannot open files, run commands or reach the internet. They are given only the code they need, return one answer, and are told to treat that code as information to describe, never as instructions to follow.
Meet all the agents →Stage 02 · write the specification
Stage 03 · add what was never written
Stage 04 · plan the next phase
Stage 02 also has 8 supporting code agents that sweep, name and check what the code reading finds.
What Glue does not do
Your code and your access
Tell us your name, your email, a phone number and your organization. Someone at Antino will look at your request and, if we give you access, set up your guest sign-in.
You can also write to hello@antino.ai.