Keep your AI coding
on course.

Running npx codekeel install sets up everything in one shot: it registers a few hook commands in your global ~/.claude/settings.json, adds /decision to ~/.claude/commands/, and seeds this project's ledger/DECISIONS.yaml with six curated starter rules covering secrets, personal information, migrations, failing tests, lint/type suppressions, and .env files. No account, no network call — all local.

The hooks and slash commands are global — set up once per machine, shared by every project you touch. The ledger isn't: it lives inside each repo, so you need to run npx codekeel install again (or just npx codekeel ledger init) in every new project to get one seeded there. Already-installed hooks are skipped automatically, so re-running anywhere is always safe.

Each starter rule is an ordinary entry once written: supersede any of them the same way you'd supersede a decision you recorded yourself.

FUNCTIONS.md and PROJECT_STATE.mdaren't part of that setup step — they show up separately, written by a hook that runs after your first real Claude Code session in the project.

Three separate layers, only one of which needs an Anthropic API key:

• 1. Scope-confirmation gate — free

  Before an edit lands, if a turn is touching an unusually broad set of files, or a file that looks auth/permissions/billing/migration-related and wasn't mentioned, Claude Code pauses and asks you first. No ledger entry required — this runs on every session.

• 2. Pattern check — free

  A mechanically-checkable decision (e.g. "migrations are append-only") gets checked instantly and locally — no network call, no account needed, and no meaningful delay.

  Checking edit against D-002 (local)...
  ✗ Blocked: modifies an existing migration

• 3. Live semantic check — needs ANTHROPIC_API_KEY

  A decision that depends on judgment (e.g. "don't change styling unless the prompt asked for it") can't be reduced to a pattern — the diff and the instruction that triggered it get checked by a live model call. This is the one part of the three that needs a key.

Only for the live checks, and only when ANTHROPIC_API_KEYis set: the file path, the diff, and the prompt that triggered it are sent straight from your machine to Anthropic's API for a verdict, then discarded. codekeel has no server in that path and never sees this traffic.

Nothing is retained anywhere by codekeel — not your code, not your diffs, not the verdicts.

The free CLI handles recording and reviewing decisions locally, and even enforces the decisions that reduce to a plain pattern (no model call, no key). Setting ANTHROPIC_API_KEY puts a live model in the loop, called directly from your machine, for the parts that genuinely need judgment:

• Live semantic enforcement on judgment-dependent decisions

  Every diff under a decision's scope that can't be reduced to a pattern gets checked against it in real time, with an explanation if it's about to drift.

  Checking edit against D-014...
  ✗ Blocked: removes Zod validation
    from POST /api/webhook

• Rejects vague rules that can't be enforced

  An admissibility check catches invariants too fuzzy to actually verify, before they clutter the ledger with rules nothing can check.

  /decision
  > Write clean code
  
  ✗ Too vague to enforce.
    Try: "No function exceeds 50 lines"

• Full project memory carried forward automatically

  PROJECT_STATE.mdgets fed back into every new session's context on its own, instead of just sitting in the repo waiting to be read.