Printing Press: CLI Factory for AI Agents
Nate Herk | AI Automationgo watch the original →
the gist
Printing Press delivers 50+ pre-built CLIs and a factory to build custom ones for sites without APIs, slashing token use vs MCPs (35x less) for Claude.
The Breakthrough
Printing Press equips Claude with a library of 50 pre-built CLIs (e.g., ESPN, Flight Goat, Movie Goat) and a factory that generates custom CLIs from any tool or site in minutes via natural language instructions.
What Actually Worked
- Users install the starter pack via
pp install starterafter setting up Go (Google's open-source language for CLI tools); Claude handles setup with provided links: https://github.com/mvanhorn/cli-printing-press, https://github.com/mvanhorn/printing-press-library, https://printingpress.dev/. - The factory builds custom CLIs through Claude prompts like "use the CLI factory to create a CLI for Hacker News to pull articles"; it researches endpoints, catalogs features, generates Go code, and verifies with dogfood runtime testing, producing commands like
./hn top --min-points 100for top stories. - Pre-built CLIs handle sites without APIs (e.g., School, Craigslist, eBay) via reverse-engineered Chrome sessions or scraping; School CLI fetches wins category posts with
school-pp wins --category wins --limit 3, returning links and summaries. - Agents invoke CLIs naturally (e.g., "what NBA games tonight?" uses
pp-espn schedule nba); output stays clean (200 tokens of text vs raw JSON). - Users package and share CLIs via private GitHub repos for teams, swapping API keys locally.
Before / After
MCP servers used 35 times more tokens than CLI on the same task. CLI achieved 100% reliability; MCP dropped to 72% as tasks got harder. School CLI request consumed 260 input + 132,000 output tokens externally, but only 2,000 entered Claude's context.
Context
Agents struggle with token costs from massive API JSON responses or MCP overhead (tool descriptions, server runtime). Printing Press creates local, SQLite-backed CLIs with lazy discovery, no rate limits on calls, and composable commands, ideal for pay-per-token setups like Claude. It targets sites lacking APIs (e.g., School, ESPN) by enabling quick CLI generation, letting agents chain tools efficiently without context pollution.
Notable Quotes
- "MCP used 35 times more tokens than the CLI on the same task and reliability drops from 100% with the CLI to 72% with MCP as tasks get harder."
- "It sent about 260 tokens to school it sent me back about 132,000 tokens but what's important is that none of that hit my context window."
- "CLI are built for agents especially when you're paying per token or you know managing your session limit."