Microsoft's dashboard for your own AI coding habits, not your code

A dashboard for how you actually use an AI coding assistant

This is a Microsoft-published VS Code extension, MIT licensed, that reads your local AI coding session logs and turns them into a dashboard of your own habits: practice scores and weekly trends, forty-five rule-based checks for specific anti-patterns across prompting, session hygiene, code review and tool use, a breakdown of how much AI-generated code you produce by language and model, discovery of repeated prompts worth turning into reusable skills, and an assessment of how healthy your workspace's own context and instruction files are.

The framing in the README is worth taking at face value: this analyses usage, it does not replace judgement. Nothing here writes code, reviews a pull request, or makes a decision for the developer. It observes what already happened in your session logs and reflects it back as measurement, on the premise that a developer who can see their own prompting habits, session length patterns and repeated mistakes is in a better position to improve them than one working entirely blind. That premise, that visibility into your own behaviour is itself a lever for improving it, is the whole thesis of the tool, and it is a reasonable one applied to a skill, using an AI coding agent well, that most developers are currently learning by feel rather than by any structured feedback loop.

Any harness, one dashboard, and what that actually requires

The tagline promises analysis across any harness through one dashboard, and the pages described later back that up concretely: a harness breakdown that distinguishes an interactive application session from a terminal-driven command-line session, and a design that reuses the identical webview bundle whether it runs inside VS Code or inside a separate application's own canvas surface.

Supporting several different tools' session-log formats under one unified analysis is a real integration cost that a single-harness tool never has to pay, and it is the correct scope for a project whose whole point is measuring a developer's overall relationship with AI-assisted coding rather than their usage of one specific product. A developer who splits their work between an editor extension and a command-line tool, which is an increasingly common pattern, gets a genuinely unified picture only if the analysis actually understands both log formats and merges them into one coherent view rather than requiring two separate dashboards that never talk to each other.

Everything stays local, which is the only credible design for this data

The README's stated architecture is that the extension reads local session logs and that no data leaves the machine, extending explicitly to the canvas mode running inside the separate Copilot application, which the README specifies parses those same local logs in process rather than through any server round trip.

That is close to the only credible design for a tool whose entire input is a developer's private coding session history: every prompt they wrote, every file they touched, every mistake the assistant made and every correction they issued. Session logs of that kind are a rich record of how someone actually works, including plenty of half-formed thoughts, wrong turns and private project details that a developer would reasonably not want leaving their own machine merely to be measured. Local-only processing removes the question of what a remote service does with that data by making the question moot, and choosing that architecture for a tool built by a large software vendor, rather than the more commercially convenient path of centralising the analysis, is worth noting explicitly as a deliberate design decision rather than an accident.

The mode you run it in changes which features you actually get

The extension supports two distinct run modes, and the README is specific and honest about how they differ rather than presenting one as a strict superset of the other.

Inside VS Code itself, the full feature set is available, including the AI-driven Skill Finder that surfaces repeated prompts worth turning into reusable skills, a Learning Center, a Level Up progression section, and an AI-assisted review of the workspace's context health. Running instead as a canvas inside the separate GitHub Copilot application reuses the identical webview bundle and processes the same local logs, but the README states plainly that those four specific features are hidden there because they depend on the local VS Code language model, which the canvas host does not provide. Everything driven purely by on-disk log data, the Dashboard, Timeline, Coding Moments, Output and the anti-pattern detection, works identically in both modes.

Naming exactly which four features are unavailable, rather than leaving a user to discover the gap by clicking around, is the right way to document a genuine capability difference between two run modes of the same tool, and it tells a prospective user precisely what trade they are making by choosing the lighter-weight canvas path over installing the full extension.

You have to build it yourself, and that is stated up front

The README states directly that the extension is not published to a marketplace or a releases page, so anyone who wants it has to build the installable package themselves from source. Two build paths are offered depending on what is already on the machine.

For anyone without a local Node.js toolchain, a dev container path uses VS Code's own container tooling with Docker or Podman:

npm ci
npm run package

For anyone with Node.js and npm already available, the equivalent local path is a straightforward clone and build:

git clone https://github.com/microsoft/ai-engineering-coach.git
cd ai-engineering-coach
npm ci
npm run package

Either path produces an installable package file, which is then installed through the command line or through the editor's own install-from-file command. Requiring a self-build rather than a one-click marketplace install is a genuine extra step for anyone evaluating this, and it is consistent with a project that reads sensitive local usage data by design: building from source you can inspect yourself is a materially different trust proposition than installing a compiled artefact from a store listing, and for a tool whose entire premise rests on a no-data-leaves-the-machine promise, being buildable and auditable from source is the more coherent choice even at the cost of installation convenience.

What to weigh before adopting it

Two features documented in the pages table are explicitly marked as partially unavailable in the current build: a token-usage breakdown in the Output page is temporarily hidden, and a monthly token-budget burndown page is temporarily disabled. Naming those limitations directly in the feature table, rather than shipping a placeholder that quietly does nothing, is a small honesty that matters for setting the right expectation before someone goes looking for a feature the README already told them is not currently working.

The project reports 4,225 stars, 581 forks and 41 open issues, and the last push was 2026-09-17, so it is an actively maintained, officially published Microsoft repository rather than a side experiment, even though its distribution model still requires a manual build.

Before adopting it, three steps in order. Pick the build path matching your setup, the dev container route if you would rather not install Node.js locally, or the direct local route if you already have it. Decide whether you want the full VS Code extension or the lighter canvas mode inside the GitHub Copilot app, understanding that the canvas path drops four AI-assisted features that depend on the local language model. And treat the forty-five anti-pattern rules and the practice-score trends as a mirror on your own habits worth reading regularly rather than a one-time report, since the entire value of a local usage dashboard comes from checking it often enough to actually change something.

Editorial conclusion

AI Engineer Coach fits a developer who wants to see their own AI-assisted coding habits laid out concretely, prompting patterns, session hygiene, repeated mistakes and how much generated code they actually ship, rather than a tool that writes or reviews code itself, and its forty-five anti-pattern rules give that feedback real specificity rather than a vague quality score. Its local-only processing, extending even into the alternative canvas mode inside the GitHub Copilot app, is the only credible architecture for a tool whose entire input is a developer's private session history, and building it from source yourself rather than installing from a marketplace listing is a reasonable extra step for exactly that kind of data. Pick the build path matching your setup, decide between the full extension and the lighter canvas mode based on whether you need its four language-model-dependent features, and treat the dashboard as something to revisit regularly rather than a one-time report, since its value comes from actually changing behaviour over time.

Frequently asked questions

Official sources

1. Issues
2. License: MIT
3. microsoft/AI-Engineering-Coach on GitHub
4. README