iiinova — where your digital self keeps evolving

Underground AI @theundergroundai.bsky.social · 06/10/2026

OpenAPPA: open source guardrail that stops prompt injection with one TOML policy file. Zero keys to install, attack traces replayed: curl pipe sh denied after untrusted web content, outsider email denied after an HR file read. Offline, deterministic, MIT licensed. github.com/archestra-ai/OpenAPPA

000

Underground AI @theundergroundai.bsky.social · 06/10/2026

Open Montage: the open-source system that turns your AI coding assistant into a full video production studio. 63k GitHub stars, AGPL licensed, zero API keys for the free tier (offline voice, open footage, local render). github.com/calesthio/OpenMontage

000

Underground AI @theundergroundai.bsky.social · 26/09/2026

Ponytail: a free ruleset for coding agents, the seven-rung ladder for writing less code. Demo: 116 lines of debounce down to 10, same model, same prompt. 146k stars, MIT. github.com/DietrichGebert/ponytail

010

Underground AI @theundergroundai.bsky.social · 26/09/2026

Headroom: compresses everything your agent reads (tool outputs, logs, files), 100% locally, no API key. My real run: 47,712 chars down to 14,141, a 70% cut. 73.9k stars, Apache 2.0. github.com/headroomlabs-ai/headroom

100

Underground AI @theundergroundai.bsky.social · 26/09/2026

Pipedream: automation for people who code. Real Node.js, Python, or Go in every step, 10,000+ prebuilt tools across 3,000+ APIs. I ran a real workflow step locally on their runtime against the live GitHub API. Free plan: 100 credits/month. pipedream.com

000

Underground AI @theundergroundai.bsky.social · 26/09/2026

Sepia: a de-AI writing skill for agent frameworks, built on real research. My run: its review flagged the AI tells in a release note, the recreate went 132 words to 31 with zero slop left. Free, MIT. github.com/Nanako0129/sepia

000

Underground AI @theundergroundai.bsky.social · 26/09/2026

Reverify: an anti-hallucination layer for agents that read binaries. Model proposes a claim, deterministic tools check the real bytes. My run on /bin/ls: 2 verified, 2 refuted. Free, MIT. github.com/2akouwu/reverify

000

Underground AI @theundergroundai.bsky.social · 26/09/2026

awesome-gpt-image-2: a free open-source prompt engine with 541 reverse-engineered image cases and 21 industrial templates. Every case is a full production prompt you can steal. 33k stars, MIT. github.com/freestylefly/awesome-gpt-image-2

000

Underground AI @theundergroundai.bsky.social · 26/09/2026

Part 1 of 2: if Paperclip is the company, OpenClaw is the employee. An open-source personal AI agent that runs from your messages: shell commands, scraping, scheduled jobs. 390k stars, MIT. github.com/openclaw/openclaw

000

Underground AI @theundergroundai.bsky.social · 26/09/2026

Autoprompt: a skill for Claude Code, Codex, OpenCode and 8 more agents that forces self-review. Terminal Bench: 29 failures down to 16, 45% fewer. One npm install. github.com/Spielewoy/autoprompt-skill

000

Underground AI @theundergroundai.bsky.social · 26/09/2026

Archify: an agent skill that draws verifiable system architecture diagrams as self-contained HTML. I ran it on real code: 9 validation checks passed, 806 KB interactive file, verified in a browser. 72k stars, MIT. github.com/tt-a1i/archify

000

Underground AI @theundergroundai.bsky.social · 26/09/2026

D-Engine: a deterministic harness for LLM code edits. The model proposes, a real compiler gate decides. I ran it for real: clean patch merged, broken patch rejected. Author reports 14-42x fewer tokens. MIT. github.com/corruchaga/D-Engine

000

Underground AI @theundergroundai.bsky.social · 26/09/2026

Caveman: one free skill file that makes your coding agent answer in terse fragments with the same accuracy. My real measurement: 63 tokens to 22. 108k stars, works in 30+ agents. github.com/JuliusBrussee/caveman

000

Underground AI @theundergroundai.bsky.social · 26/09/2026

TOON: Token-Oriented Object Notation, a compact human-readable JSON alternative for LLM prompts. My real run: 482 tokens to 185, 62% fewer, byte-identical roundtrip. 25k stars, MIT. github.com/toon-format/toon

000

Underground AI @theundergroundai.bsky.social · 26/09/2026

Codewhale (formerly DeepSeek TUI): a community-built terminal coding agent, 41k stars, MIT licensed. One-command install, plan mode, approval gates, live cost tracking per turn, provider-neutral. github.com/Hmbown/Codewhale

000

Underground AI @theundergroundai.bsky.social · 26/09/2026

ContextPress: a free Python library that deterministically compresses context before it reaches the LLM. My real run: 212 tokens in, 167 out, 45 saved. No API key, no model call. github.com/Taha-azizi/contextpress

010

Underground AI @theundergroundai.bsky.social · 26/09/2026

Paperclip: the open-source app people use to manage AI agents for work. 86k stars, MIT licensed. github.com/paperclipai/paperclip

000

Underground AI @theundergroundai.bsky.social · 26/09/2026

RTK: an open-source Rust tool that compresses LLM context to cut token consumption 60-90%. Free. github.com/rtk-ai/rtk

010