CLI Printing Press generates command-line tools for AI agents from an API, an app, or a website. For each target it produces a Go CLI, an MCP server sharing the same client, local store, and auth, and a Claude Code skill. It works from an OpenAPI spec, a HAR capture, or just a URL: for sites with no published API, a browser-sniff step captures traffic and reverse-engineers the spec.
The workflow runs as a set of agent skills (/printing-press ) that drive a generator binary. A research phase studies the official API docs and existing community CLIs and MCP servers, an absorb gate catalogs every feature those tools offer, and generation builds all of them plus compound commands that only work with local data. A shipcheck phase then verifies the result, and an optional read-only smoke test runs against the live API. Claude Code is the tested host, and the skills also install into Codex. A Codex mode delegates the code-writing phase to Codex CLI while Claude keeps research, planning, and review.
Generated CLIs mirror high-volume resources into domain-specific SQLite tables with FTS5 search and incremental sync, which enables offline search, raw SQL, and cross-resource commands such as stale, bottleneck, and reconcile. A profiler classifies each API into an archetype (project management, communication, payments, infrastructure, content) and adds matching workflow and insight commands. Output is agent-shaped: JSON when piped, a compact mode that keeps only key fields, typed exit codes, and errors that name the wrong flag and the correct usage.
Every CLI is graded on a 100-point, two-tier scorecard covering infrastructure patterns and domain correctness (path validity, auth protocol, sync correctness, type fidelity, dead code). Published CLIs go to a public library with a provenance manifest, research notes, and verification proofs. Generated CLIs cover the agent power-user surface rather than every option of a vendor's official CLI, and the project notes that many services prohibit automated access in their terms.
Features
- Three input modes: an OpenAPI or GraphQL spec, a HAR file, or a URL that the browser-sniff gate turns into a spec
- Dual interface: each API gets a Cobra CLI and an MCP server built from the same client, store, and auth
- Ecosystem absorb: catalogs features from competing CLIs, MCP servers, and skills and builds them before novel commands
- Local data layer: domain-specific SQLite tables, FTS5 search, cursor-based incremental sync, and a raw SQL command
- Insight commands: archetype-driven commands such as health, similar, trends, bottleneck, forecast, and reconcile
- Agent-native output: auto-JSON when piped, field selection, a compact mode cutting 60–80% of tokens, dry runs, and typed exit codes
- Scorecard: a 100-point grade split between infrastructure patterns and domain correctness, with anti-gaming rules
- Verification: structural dogfood checks, proof of behavior, and a read-only live API smoke test
- Codex mode: offloads code generation to Codex CLI and falls back to local generation after three failures
- MCP intents: declarative multi-step MCP tools composed from endpoints, with optional HTTP transport
- Library workflow: polish, publish, and amend skills turn a CLI or a dogfood session's friction into a library pull request
