Skip to content

AI adoption, measured with its guardrails

Know how far AI adoption has really got, repository by repository.

Engineering teams adopt AI tooling in fragments: good ideas spread engineer by engineer, but the setup is rarely consistent. Chamlis reads your GitHub organization, places every repository on one fixed ladder, and shows each one its single next step.

Request a demoSign in with GitHub

Sign-in is limited to members of an organization where Chamlis is installed.

Working on your own? Chamlis is set up for organizations today. Tell us you want it for personal repositories and we will write when it opens.

Example scan · acme-dev
RepositoryLevelGuardrailsNext step
payments-apiL24/4at the highest scored level
web-checkoutL13/4next: require status checks
data-pipelineL14/4next: commit REVIEW.md
legacy-billingL01/4next: commit AGENTS.md
  • Read-onlyA GitHub App that cannot write to your repositories.
  • Evidence-basedLevels come from committed files and enforced settings, not surveys.
  • ConservativeMissing or ambiguous evidence fails the check.

Who it is for

Useful to the person leading fifty engineers, and to the person working alone

The ladder is the same at every size. What changes is the question you bring to it.

  • For engineering leaders

    One comparable picture across every team, without a survey.

    • The organization overview shows how many repositories sit at each level and how that moves over time; team roll-ups show the same counts for each team.
    • Guardrails are reported next to adoption, so faster never hides less safe.
    • The evidence is committed files and enforced settings, which you can show to security and audit.
  • For engineers

    A short, checkable list for your repository. Nothing about you is ranked.

    • Each repository shows its single next step, and every check is something you can verify in the repository yourself.
    • Your own activity signals are visible to you and the maintainers of your teams, and to no one else.
    • No prompts, transcripts or keystrokes are collected, and there is no leaderboard.
  • For solo maintainers

    The same ladder works for an organization of one.

    • Install it on your own GitHub organization and see every repository you maintain on one page.
    • It works as the setup checklist nobody writes for themselves: agent instructions, review, branch protection.
    • When there is no second pair of eyes, it shows which guardrails are standing in for one, and which are missing.

Why Chamlis

A shared AI setup, without a project to get there

Most teams already use AI coding tools. What is missing is a setup the whole team shares, and review that keeps pace with it. Chamlis shows the gap and closes it one small step at a time.

Easy to adopt

  • One install, nothing to instrumentA read-only GitHub App for the organization. Nothing goes on laptops and nothing is added to CI.
  • No surveys, no telemetry rolloutLevels come from files, branch rules and pull-request history already in GitHub. Nobody fills in a self-assessment.
  • No rubric to writeChamlis ships one standard: agent instructions, an MCP config, review rules and four guardrails. There is nothing to choose or configure first.
  • One next step per repositoryEach repository shows a single next step, usually one small file or one branch setting. Not a report.

Built for small teams

  • No platform team neededLarger companies give this job to a platform team. Chamlis is the standard and the checklist that team would have written.
  • No minimum sizeScoring is per repository. An organization with one repository and one engineer gets the same ladder, with no developer portal to set up first.
  • Your setup stops living in your headAgent instructions, MCP servers and review rules are committed to the repository. Every agent session starts from them, and so does a contractor or a first hire.
  • Nothing to redo as you growNew repositories and people are picked up by the nightly scan and scored on the same ladder.

One limit for a team of one: Level 2 requires a human approval on every pull request, which needs a second person. A one-person repository can reach Level 1 today.

How it works

From install to a clear picture, with nothing to instrument

  1. Install the read-only GitHub App

    One installation for the organization. It can read repository contents and settings; it cannot write.

  2. Every repository is scanned

    Chamlis collects evidence from committed files, branch protection and pull-request history, every night.

  3. See where each one stands

    Organization, team and repository views show the level reached, the guardrails in place, and the next step.

The ladder

One ladder, small steps

A repository's level is the highest one it meets together with every level below it. Each level is a short list of things you can check in the repository yourself.

Levels 0 to 2 are scored today. Levels 3 to 5 are shown, and not scored yet.

  1. L0
    Level 0 · Ad hocNone of the steps below.
  2. L1
    Level 1 · BaselineAgent instructions (AGENTS.md or CLAUDE.md) with real content, and .mcp.json, committed.
  3. L2
    Level 2 · ReviewedAn AI reviewer running on pull requests (with REVIEW.md where it is Claude Code Review), and every guardrail in place.
  4. L3
    Level 3 · SharedNot yet assessedA shared standards plugin installed, with no drift from the organization's standard.
  5. L4
    Level 4 · Spec-drivenNot yet assessedMerged feature pull requests link to a spec.
  6. L5
    Level 5 · Full loopNot yet assessedDecisions and sign-off are traceable from spec to merged pull request.

Guardrails

Guardrails measured alongside

Speed does not count as progress on its own. A repository with a guardrail missing is never reported above Level 1, whatever else it has.

  • A CODEOWNERS fileEvery change has a named owner to review it.
  • Required human approval on pull requestsA person signs off before anything merges.
  • Required status checksTests and checks must pass, enforced by branch protection.
  • An AI reviewer on every pull requestClaude Code Review by default. Another AI reviewer counts, and is reported as non-standard.

When evidence is missing or ambiguous, the check fails: Chamlis under-reports rather than give a false pass.

Playbooks

A library of proven ways to run the work

In design

The ladder gets a repository ready for agents. Playbooks are what you do next: step-by-step processes for the situations engineering teams meet, each drawn from what a real team did, that Claude Code follows with your people at the points that need them.

  • Platform unificationTwo or more platforms must become one.
  • Codebase mappingYou need to understand a codebase you did not write.
  • Legacy modernizationAn old system must move to a new stack.
  • Large-scale migrationThe same change is needed across hundreds of files or repositories.
  • New feature in an existing productSpec, research, plan, then build phase by phase.
  • Greenfield serviceA new service or product from an empty repository.
  • Prototype / spikeA quick build to answer a question.
  • Incident responseProduction is broken.
  • Bug investigationA bug with an unknown cause.
  • Security review and remediationReviewing changes for security, and fixing what is found.
  • Test backfillA codebase with too few tests.
  • Onboarding and agent docsA new person or team joins, or agent docs need upkeep.

How a playbook runs

  • Start with one sentenceTell Claude Code what needs doing. It picks the matching playbook and works through its steps, each marked for the agent, a person, or both.
  • Pick up where the last session stoppedProgress is kept as files in a repository you own, so any engineer can continue in a new session without a briefing.
  • Questions go to the people who knowWhen the agent cannot settle something it takes a sensible default, carries on, and opens a GitHub issue for the engineer or product manager who can answer.
  • Sign-off is a pull request approvalThe work stops at each sign-off until the named owner approves, using the review rules you already have.

Who it is for

  • An engineering lead with work ahead that is too big for one session and touches several repositories or teams.
  • A team already using Claude Code day to day, where what is missing is a process rather than a tool.
  • Work where no one person holds the whole picture, so engineers and product managers each need to add what they know.

The playbooks are written. Running one from Claude Code, and the page that tracks it, are designed and not built yet. The work happens in your engineers' own sessions and in your own repository; Chamlis only reads it, to show the step, the open questions and the pending sign-offs on one page.

Beyond engineering

One process for product, design and engineering

In design

Agents speed up writing code. Deciding what to build and how it should work is still done the old way, and it happens somewhere the agent never sees. Chamlis playbooks bring requirements and design into the same process as the code, for a team of three functions or for one person doing all of it.

  • Steps assigned to rolesEach step of a playbook says who acts: product, design, engineering or the agent. A feature goes from brief to flows to spec to build in one file in your repository, and each handoff is a pull request someone approves.
  • Working alone, you hold every roleThe agent takes the other side of each handoff. It questions the brief the way an engineer would and walks the flow the way a new user would, so the product and design steps still happen.
  • A better process for a project that already worksPlaybooks for adding end-to-end tests, and for turning production errors into fix pull requests that a person reviews.
  • The next step becomes a playbookChamlis reads the repository, recommends the playbook that fits what it finds, says why, and sees at the next scan that it was adopted.

This is designed and not built yet. It is for solo maintainers, small product teams, and engineering and product leaders who want one process. There is no score for a person or a function, your briefs and designs stay in your repository and your own tools, and Chamlis still only reads GitHub.

Impact analysis

Is AI adoption paying off?

In design

Impact analysis uses your organization's merged pull-request history, the same read-only GitHub data Chamlis already scans, to answer one question: did AI adoption make delivery faster without making it riskier? It compares pull requests co-authored with an AI coding agent against the rest, after adjusting for the size of the change, the type of change and the repository, and reports a speed figure next to risk figures, each with a plausible range.

For: A VP of Engineering or CTO at an organization with more than 10 engineers who has to show that AI adoption is paying off without eroding quality. Organization and team level only, never per person. Part of the paid plan: what one analysis found, and what it uses.

Trust

What it does not do

A measurement tool only helps if engineers and security teams are both comfortable with it running.

  • It never writes to GitHub.The GitHub App is read-only: it requests no write permission, and Chamlis makes only read requests.
  • It is not surveillance.No prompts, session transcripts or keystroke data leave anyone's machine. Per-person signals come only from git and pull-request history.
  • It does not rank people.There is no leaderboard and no per-person metrics table.
  • Your data stays with your organization.People sign in with GitHub, and only members of your organization can see its dashboard.

FAQ

Questions and answers

What is Chamlis?

Chamlis is a read-only GitHub App and dashboard, built by Lighty AI, that scores every repository in a GitHub organization on a fixed six-level AI-adoption ladder and checks four security and quality guardrails alongside.

How does Chamlis measure AI adoption?

From evidence in the repositories themselves: committed files such as AGENTS.md, CLAUDE.md, .mcp.json and REVIEW.md, branch-protection settings, and pull-request history. Every repository is scanned every night. Nothing is self-reported, and missing or ambiguous evidence fails the check.

What are the levels of the AI-adoption ladder?

There are six: Level 0 Ad hoc, Level 1 Baseline, Level 2 Reviewed, Level 3 Shared, Level 4 Spec-driven, Level 5 Full loop. A repository's level is the highest one it meets together with every level below it. Levels 0 to 2 are scored today; levels 3 to 5 are shown and not scored yet.

Which guardrails does Chamlis check?

Four: a CODEOWNERS file, required human approval on pull requests, required status checks, and an AI reviewer on every pull request (Claude Code Review by default). A repository with a guardrail missing is never reported above Level 1.

Does Chamlis need write access to GitHub?

No. The GitHub App is read-only: it requests no write permission, and Chamlis makes only read requests.

Does Chamlis monitor or rank individual engineers?

No. No prompts, session transcripts or keystroke data leave anyone's machine, per-person signals come only from git and pull-request history, and there is no leaderboard and no per-person metrics table.

Who can see an organization's results?

Only members of that GitHub organization. People sign in with GitHub, and the dashboard is limited to members of the organization where Chamlis is installed.

How do we start using Chamlis?

Request a demo at https://lighty.ai/request-demo or email contact@lighty.ai. Setup is one installation of the read-only GitHub App for the organization, with nothing to instrument in the repositories.

See it on your own organization

Whether you run an engineering organization or a handful of your own repositories, tell us about it and we will walk you through Chamlis and what an installation involves.