CLI Printing Press generates a command-line interface and a matching MCP server for an API or a website, designed for use by AI coding agents. It reads the official API documentation, catalogues the features of existing community CLIs, MCP servers, and skills for the same API, and can capture browser traffic to reverse-engineer an API that has no published spec. The output is a Go CLI built on Cobra plus an MCP server that share one client, one local store, and one auth layer.
The generator is a Go binary driven by a set of skills for Claude Code (/printing-press ), and can also be set up for Codex and Cursor. A run moves through phases: resolve the spec or URL, write a research brief, catalogue the ecosystem into an absorb manifest, optionally sniff a website, generate the wrapper, build the data layer and compound commands, then verify. Three entry paths are accepted: an OpenAPI spec (--spec), a website URL, or a HAR file from browser DevTools (--har). REST and GraphQL specs are both supported.
A central design idea is a local-first data layer. High-gravity resources get domain-specific SQLite tables with FTS5 full-text search and incremental sync with cursor tracking, so sync pulls data down, search queries it locally, and sql exposes raw queries. Because the data lives in SQLite, the generator also emits compound commands that join across resources and history, such as stale, orphans, health, bottleneck, and reconcile. The profiler classifies each API into a domain archetype (project management, communication, payments, infrastructure, content) that selects which workflow and insight commands are generated.
Generated CLIs follow a fixed agent-facing contract: table output in a terminal, automatic JSON when piped, a --compact mode that returns only high-gravity fields, typed exit codes (0 success, 2 usage, 3 not found, 4 auth, 5 API, 7 rate limited), and --dry-run for safe exploration. Every command carries flags such as --json, --select, --stdin, --csv, --quiet, --yes, --no-input, and --no-cache.
Before a CLI ships it passes four mechanical checks: a two-tier scorecard, a structural dogfood pass, proof-of-behavior checks, and an optional read-only live smoke test when an API key is present. Finished CLIs can be published to a shared library, polished with a fix-up pass, or amended from a Claude Code session transcript.
Features
- Dual output: one spec produces a Cobra CLI ( -pp-cli) and an MCP server ( -pp-mcp) that share the same client, store, and auth
- Three input paths: an OpenAPI spec, a website URL captured through a browser-sniff gate, or a HAR file
- Ecosystem absorb gate: catalogues features from existing plugins, MCP servers, skills, and CLIs for the API into a manifest the generated CLI must match
- Local SQLite data layer: domain-specific tables, FTS5 search, incremental sync with cursors, and a raw sql command
- Compound commands: archetype-driven workflow and insight commands such as stale, health, similar, trends, and reconcile
- Agent-native flags: auto-JSON when piped, --compact, --dry-run, --select, and typed exit codes on every command
- Two-tier scorecard: 100-point scoring split between infrastructure patterns and domain correctness, with anti-gaming rules
- Proof of behavior: path, flag, pipeline, and auth proofs with automatic removal of dead code
- Live smoke test: GET-only checks against the real API with --limit 1, a 10s timeout, and a stop on 401
- Codex mode: delegates code-generation tasks to Codex while Claude handles research, planning, and review
- Library workflow: publish, polish, and amend commands plus an auth doctor for each printed CLI's env vars
