The fitness tracker for engineering skill

co-pilot's×co-pilot=co²

You already build with AI every day. CoSquared analyzes your Claude Code, Cursor, and Codex sessions — locally, on your machine — and coaches you to get measurably better at it.

your terminal — 2 minutes to your Builder Profile

$ curl -fsSL https://cosquared.astar.inc/analyze.sh | bash

Free · raw code and transcripts never leave your machine · delete everything with one command

Watch your habits form.

Every sync updates your Builder Profile. Weekly reports call out what improved, what slipped, and the one habit to focus on next.

cosquared — dashboard

week 1

AI Collab

Testing

Shipping

Testing discipline · 8 weeks

↑ Improved

Tests ran in 2 of 6 sessions, up from 1/6.

How it works

Two minutes. One paste.

01
Paste one command
Installs the cosq CLI and runs your first analysis. One browser tab opens to link your email — that's the entire signup.
02
Analysis runs on your machine
Your AI sessions and git history are parsed locally; secrets are redacted before anything is serialized. Run cosq inspect to see the exact bytes before they ever move.
03
Get coached, weekly
A Builder Profile with three scores, then a weekly report with one challenge you can actually check the next week.

your terminal — the entire onboarding

The scores

Three scores. No composite.

Deterministic rules over your own sessions and commits — not an LLM's opinion of you. A single “engineering score” would be a lie, so there isn't one.

AI Collaboration

How well you drive the machine.

How you work with AI — do you give it clear instructions, and do you check that its code actually works before accepting it?

Testing Discipline

Whether tests run before you ship.

Did tests really run while you built? We count the coding sessions where they did — actual runs, not good intentions.

Shipping Momentum

Whether the work lands.

Does the work you start actually get finished? We check whether your coding sessions end in real commits, from git history.

Every score cites the sessions and commits it came from. Thin evidence gets labeled low confidence — never inflated. Read the published methodology →

Privacy, by construction

Your code never leaves your machine.

Raw code never leaves your machine

Code, transcripts, credentials, env values — never uploaded. Redaction runs before anything reaches the network path; uploads are metadata-level evidence only.

Never visible to managers

Your scores are for you. Individual profiles can't be surfaced to a manager or admin — a product invariant, not a settings toggle.

Open-source collector

Every line that touches your data is public and auditable — collectors, redaction, and the full scoring methodology.

See the exact bytes first

cosq inspect prints output byte-identical to what cosq analyze uploads. Not a summary of it — the same bytes.

Delete everything, one command

cosq delete hard-deletes your server-side data and hands back a deletion receipt.

Read the full privacy doc →

Every invariant above, in writing — the same doc that ships in the open-source repo.

Evaluation API · design partner preview

Build your own product on evidence-linked engineering signals.

Submit strict, free-text-free workflow events and get deterministic dimension results with confidence, coverage, and evidence references. Test keys are self-serve; live access is manually approved during the design-partner phase.

Read the docs Sign in for a test key →

one request — versioned results

$ curl -X POST \
  https://cosquared.astar.inc/eval/v1/evaluations \
  -H "Authorization: Bearer $COSQUARED_EVAL_KEY" \
  -d @evaluation.json

HTTP/2 201
{
  "status": "completed",
  "events": { "accepted": 13, "duplicates": 0 },
  "results": [{
    "profile": "ai_collaboration_v1",
    "dimensions": [{
      "key": "task_framing",
      "score": 74,
      "confidence": "normal"
    }]
  }],
  "limitations": [
    "methodology_not_validated_for_consequential_decisions"
  ]
}

The coaching loop

One email a week. Receipts included.

Every report issues one checkable challenge; the next report checks it against your actual sessions. No dashboard guilt — the coach follows up.

inbox

week 8

co²

CoSquared Coach <coach@cosquared.astar.inc>

to you

Mon 7:02 AM

Your week 8 report — tests ran in 5 of 6 sessions

Last week's challenge

↑ Improved

Tests ran in 5 of 6 coding sessions — up from 2/6 last week.

AI Collaboration74 ▲+3

Testing Discipline61 ▲+5

Shipping Momentum82 ▲+2

Focus this week: keep tests running on every new endpoint — same habit, one more week.

one email a week · unsubscribe is one click, honored instantly

every number is computed locally and template-rendered — the model never invents a figure

See how you really build.