cliwright

A wright of CLIs. Forge a complete, production-grade, agent-ready command-line tool for any REST API — and loop until it provably meets every quality bar.

cliwright is a spec-gated CLI factory. Point it at an API; it researches the docs, scaffolds a Go + Cobra + GoReleaser binary to a single high standard (keyring auth, named profiles, resilient client, an MCP server, an agent guard, CI/CD, signed releases, packaging), and drives the build to a deterministic acceptance gate so "done" means provably done, not asserted done.

It is not a new agent loop — it rides the runtime's native /goal loop (Claude Code and Codex both ship it) and supplies the two things a loop needs to finish honestly: a complete spec and a machine-checkable gate the completion promise is bound to.

What you get

One standard, every API. A generic typed core + thin per-resource files; full CRUD, pagination, idempotent retry, adaptive rate limiting, --dry-run curl, Ctrl-C cancel.
Secure by construction. Tokens in the OS keyring, secret flags never exposed to the MCP surface, path-confined file reads, CSV formula-injection guards.
Agent-ready. Every command annotated read-only/write/destructive; an mcp server and an agent guard derived from the live command tree.
Shipped completely. Homebrew/Scoop + deb/rpm/apk, cosign + SBOM, generated docs site, and — optionally — its own Claude Code plugin + cross-tool agent skill.

The full, self-contained playbook is GOAL.md.

Install

As a Claude Code plugin

/plugin marketplace add jjuanrivvera/cliwright
/plugin install cliwright@cliwright

As a cross-tool agent skill (Cursor, Codex, Gemini CLI, …)

npx skills add jjuanrivvera/cliwright

Use

# Slash command (Claude Code)
/cliwright:build-cli Stripe https://docs.stripe.com/api github.com/me/stripe-cli

# Or paste GOAL.md after /goal, fill the TARGET API block, and let the loop run.

The build is complete when the acceptance gate is green:

make verify   # make check + coverage + spec-check + spec-completeness + Definition-of-Done checklist + judge rubric

How it works

Layer	Responsibility
`/goal` (native)	The loop engine — re-feeds the goal until completion
`cliwright` skill / `GOAL.md`	The spec: standard, architecture, phases, determinism rules
`make verify`	The deterministic gate the completion promise binds to
`cli-builder` agent	Runs the long multi-phase build in its own context window

Development

The plugin manifests (.claude-plugin/plugin.json, .claude-plugin/marketplace.json) are validated against the vendored schemastore schemas in .claude-plugin/schemas/, plus structural checks that every referenced component resolves and carries the frontmatter Claude Code needs to load it.

python3 scripts/validate-plugin.py   # run the same check CI runs
./scripts/install-hooks.sh           # enable the pre-commit hook (sets core.hooksPath)

CI runs the validator on every push and PR (.github/workflows/validate.yml). The full schema layer needs jsonschema (pip install jsonschema); without it the hook still runs the structural checks and CI enforces the rest. To refresh the vendored schemas, re-download claude-code-plugin-manifest.json and claude-code-marketplace.json into .claude-plugin/schemas/.

License

MIT — see LICENSE.

Cliwright

README