Mario Zechner is the creator of pi, the minimal self-extensible coding agent he took with him to Earendil in April 2026, and a two-decade open-source game-tools developer who writes long, argumentative technical posts at mariozechner.at. Facts below verified as of 2026-09-24.
He is the sharpest contrarian voice in harness design, arguing in public and in code that most agent features are baggage rather than progress, which makes his posts required reading for harness builders and his governance style the part you accept or reject.
What it is #
An Austrian (Graz-based) developer, coach, and angel investor who created libgdx, the most used Java game framework of the Android era (25.4k stars on GitHub as of 2026-09-24), handed it to contributors in 2016, and lived through the RoboVM acquisition that turned closed-source under Xamarin. His current work is pi, formerly badlogic/pi-mono, now earendil-works/pi: an MIT-licensed TypeScript toolkit (unified LLM API, agent loop, TUI, coding agent CLI) whose flagship philosophy is a system prompt under 1,000 tokens, four tools, YOLO-by-default permissions, and deliberate refusal to ship MCP, plan mode, todos, or subagents. He documents every refusal in long posts with benchmarks, most notably the November 2025 “What I learned building an opinionated and minimal coding agent” and the November 2025 anti-MCP essay showing a 225-token CLI toolset replacing a 13.7k-token MCP server. Since April 2026 pi is owned by Earendil with him at the helm, planned as three tiers: MIT core forever, Fair Source value-adds, proprietary enterprise features not yet built.
Status #
Active, well-resourced, and newly commercial as of 2026-09-24. The pi repository shows about 109k stars and 13.8k forks as of 2026-09-24, with roughly weekly releases (v0.87.1 on 2026-09-22 per our Harnesses note), and OpenClaw, the breakout assistant runtime, is built on it, which is exactly the attention he says drove acquisition offers and 3-5 calls a day before he sold to Earendil. His blog slowed after the 2025 agent-post burst (eight agent-era posts between June and November 2025, then three in 2026 through the May 30 robot post), so the repository and release notes now carry more signal than the blog. The Earendil arrangement is the open question: the MIT core is pledged as non-negotiable, but tiers two and three existed only as promises when he wrote about them in April 2026.
Strengths #
- He argues every design refusal in the open with evidence: the Terminal-Bench 2.0 submission backing his small-prompt claim, the token accounting against MCP servers, the observability argument against black-box subagents.
- His practical artifacts (the MCP-vs-CLI benchmark, the cchistory system-prompt tracker, the agent-tools CLI collection) are directly reusable by builders on any harness.
- Two decades of open-source scar tissue, especially the RoboVM closure, gives his writing on commercialization, forks, and governance a depth almost nobody else in this space has.
- He is consistently entertaining and specific, from naming pi “entirely un-Google-able” on purpose to closing as “Pidalf”, which keeps dense technical argument readable.
Cautions #
- He is an openly dictatorial maintainer: contributor issues and pull requests are auto-closed by default, and he wrote in April 2026 that he still does not trust outside contributions because everyone “is just slinging their clanker”.
- The YOLO-by-default security posture is the most contested thing he ships; the Hacker News thread on his pi post became a long sandboxing argument, with commenters countering that data can be sandboxed and that anything running outside the sandbox executes with your full permissions.
- Personal-blog cadence has dropped to a few posts a year in 2026, so there is little new argument to follow beyond release notes.
- pi’s independence ended in April 2026, and readers must now weigh pledged MIT-forever commitments against Earendil’s unbuilt commercial tiers, a bet he himself frames through his RoboVM experience.
Compared to #
- Armin Ronacher: his Earendil colleague and sparring partner, and the more measured writer; Ronacher’s “Armin is wrong” antagonist is the one to pick for architecture essays, Zechner for full-throated arguments with benchmarks.
- Boris Cherny: the maximalist platform creator versus the minimalist harness creator; read them against each other to see the whole design space from 33k-token baselines down to a sub-1,000-token prompt.
- Simon Willison: Willison catalogs what exists and names what is dangerous, Zechner tells you most of what exists should not; choose Willison for landscape coverage, Zechner for design conviction.
Bottom line #
Recommended for harness builders, capability minimalists, and anyone deciding what an agent should refuse to ship with. Not for teams that want turnkey guardrails, MCP-centric stacks, or a maintainer who accepts outside direction.
Top 5 recommended reading #
What I learned building an opinionated and minimal coding agent - the pi manifesto: a sub-1,000-token system prompt, four tools, YOLO defaults, and every refusal argued with evidence including a Terminal-Bench run.
Changes #
- 2026-09-24 - Created.
- 2026-09-24 - Added the Top 5 recommended reading section.
See also #
- Pi - the harness he builds, with star counts, release cadence, and the oh-my-pi fork that answers his refusals in running code
- Armin Ronacher - his Earendil collaborator, second-largest pi contributor, and recurring debate opponent
- Claude Code - the maximalist platform he patches, benchmarks against, and cites as the reason he built pi
- Simon Willison - the security-focused counterpoint he engages directly on prompt injection and the dual-LLM pattern
References #
https://github.com/badlogic - his GitHub profile as of 2026-09-24: 7.7k followers, pinned pi and libgdx repositories
https://hn.algolia.com/api/v1/items/46844822 - the Hacker News thread on his pi post, dominated by the sandboxing dispute