# Pi (agent framework)

> Source: https://aiwiki.ai/wiki/pi_agent
> Summary: Not to be confused with Pi, the consumer chatbot from Inflection AI, or with π0, the robot foundation model family from Physical Intelligence.
> Updated: 2026-08-07
> Fact-checked: 2026-08-07
> Categories: AI Agents, AI Code Generation, AI Tools & Products, Developer Tools, Open Source AI
> License: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/) - attribute to "AI Wiki (aiwiki.ai)"
> Cite as: AI Wiki. "Pi (agent framework)." aiwiki.ai, 7 Aug 2026. https://aiwiki.ai/wiki/pi_agent
> From AI Wiki (https://aiwiki.ai), the free encyclopedia of artificial intelligence. Reuse freely with attribution.

| Field | Value |
|---|---|
| Type | [Coding agent](https://aiwiki.ai/wiki/coding_agent) and [agent harness](https://aiwiki.ai/wiki/agent_harness) |
| Original author | Mario Zechner |
| Developer | Earendil Inc. |
| Initial development | August 2025 |
| Latest release | v0.84.1 (August 7, 2026) |
| Written in | [TypeScript](https://aiwiki.ai/wiki/typescript) |
| License | MIT (core) |
| Repository | github.com/earendil-works/pi |
| Website | pi.dev |

*Not to be confused with Pi, the consumer chatbot from [Inflection AI](https://aiwiki.ai/wiki/inflection_ai), or with [π0](https://aiwiki.ai/wiki/pi_zero), the robot foundation model family from [Physical Intelligence](https://aiwiki.ai/wiki/physical_intelligence).*

**Pi** is an open-source [coding agent](https://aiwiki.ai/wiki/coding_agent) and agent harness written in [TypeScript](https://aiwiki.ai/wiki/typescript), created in 2025 by the Austrian developer Mario Zechner and owned since April 2026 by Earendil Inc., a public benefit corporation founded by Flask creator Armin Ronacher and Colin Daymond Hanna [1][5]. Pi is built around deliberate minimalism: its default configuration gives the model only four tools (read, write, edit, bash), and its system prompt and tool definitions together fit in under 1,000 tokens, with all further capability added through a self-extension mechanism rather than built-in features [2][4]. The project is MIT licensed and had about 85,000 GitHub stars and more than 10,000 forks as of August 8, 2026 [1]. It underpins several other agent products, including Peter Steinberger's [OpenClaw](https://aiwiki.ai/wiki/openclaw) and [MiniMax](https://aiwiki.ai/wiki/minimax)'s [MiniMax Code](https://aiwiki.ai/wiki/minimax_code) [4][13][14].

## Origins

Zechner, previously known as the creator of the libGDX game development framework, began the project in August 2025 as the GitHub repository badlogic/pi-mono [1][6]. The first npm release of the pi-ai library appeared on August 30, 2025, and the coding agent CLI was first published on November 12, 2025, under the @mariozechner scope [3][17].

He described the motivation in a November 30, 2025 write-up, "What I learned building an opinionated and minimal coding agent": existing harnesses such as [Claude Code](https://aiwiki.ai/wiki/claude_code) had accumulated features he did not use, changed their system prompts between releases in ways that altered model behavior, and injected context "behind your back" without surfacing it in the interface [2]. He wanted exact control over what enters the model's context, full observability of every interaction, a documented session format, and clean support for self-hosted models. He also joked that he gave the agent "a name that's entirely un-Google-able, so there will never be any users" [2]. A February 2026 Hacker News thread on the post drew over 400 points [19].

Pi's early visibility came largely through OpenClaw, Peter Steinberger's agent project (earlier known as Clawdbot and Moltbot), which went viral in January 2026 and uses Pi as its underlying coding agent. Ronacher publicized the connection in a January 31, 2026 post, "Pi: The Minimal Agent Within OpenClaw", in which he called Pi the coding agent he used "almost exclusively" [4]. Zechner, Ronacher, and Steinberger, all based in or around Vienna, had been collaborating on agent experiments since mid-2025, a scene onlookers dubbed the "Vienna School of Agentic Coding" [6]. According to Zechner, OpenClaw's popularity brought a wave of term sheets and job offers from venture firms and large companies [6].

## Acquisition by Earendil

On April 8, 2026, Earendil Inc. announced it had acquired the Pi project and that Zechner had joined the company as a team member and "major Earendil stakeholder" [5]. Earendil is a public benefit corporation founded in 2025 by Armin Ronacher, the creator of the Flask web framework, and Colin Daymond Hanna [5][7]. Its disclosed early backers include [Accel](https://aiwiki.ai/wiki/accel) (Daniel Levine), Balderton (Daniel Waterhouse), and the founders of [n8n](https://aiwiki.ai/wiki/n8n) (Jan Oberhauser), OpenClaw (Peter Steinberger), Revolut (Vlad Yatsenko), Sentry (David Cramer), and [Slack](https://aiwiki.ai/wiki/slack) (Cal Henderson); funding amounts were not disclosed [5][7]. Zechner had himself been a small early investor in the company before the acquisition [6].

In a blog post titled "I've sold out", Zechner explained that he chose a sale to friends over founding his own startup, citing his earlier experience at RoboVM, an open-source project that was closed-sourced after its acquisition by Xamarin and then shut down by Microsoft, and his unwillingness to trade family time for a CEO role [6]. Under the deal, the repository moved from badlogic/pi-mono to earendil-works/pi, the npm packages moved to the @earendil-works scope (first release under the new scope on May 7, 2026), and pi.dev remained the project home [6][17]. Zechner retains control of technical direction and merge decisions [6].

The same announcement introduced Lefos, Earendil's first product, an assistant that communicates with users over email; it entered public alpha the same day [5]. Earendil's products are built on top of Pi, and the company says they are also built with Pi as the development tool [8].

Earendil's licensing plan for Pi was published in advance as RFC 0015 (March 30, 2026), authored by Ronacher. It commits the core to remaining MIT licensed ("Pi remains MIT licensed and that will not change"), while future value-added commercial features may be Fair Source licensed with Delayed Open Source Publication, and some enterprise features and cloud services will be proprietary [8]. The Pi name and logo are trademarked by Earendil, which the company describes as its main protection mechanism, comparing the approach to Mozilla and Linux [6][8].

## Design

Pi is developed as a monorepo of npm packages [1]:

| Package | Purpose |
|---|---|
| @earendil-works/pi-ai | Unified multi-provider LLM API ([Anthropic](https://aiwiki.ai/wiki/anthropic), [OpenAI](https://aiwiki.ai/wiki/openai), Google, xAI, Groq, Cerebras, OpenRouter, and OpenAI-compatible endpoints) with streaming, tool calling, cross-provider context handoff, and cost tracking [1][2] |
| @earendil-works/pi-agent-core | Agent runtime handling the tool-execution loop, validation, and event streaming [1][2] |
| @earendil-works/pi-tui | Terminal UI library with differential rendering that appends to the terminal scrollback rather than taking over the screen [1][2] |
| @earendil-works/pi-coding-agent | The interactive coding agent CLI that ties the packages together [1][2] |
| @earendil-works/pi-telemetry | Vendor-neutral telemetry contracts and reference adapter [1] |

The coding agent ships with a system prompt of roughly a dozen lines plus the user's AGENTS.md files, and four default tools: read, write, edit, and bash. Additional read-only tools (grep, find, ls) exist but are disabled by default [2][4]. Zechner argues that frontier models are so heavily trained on agentic coding that long harness prompts add little, a position he supported with a five-trial [Terminal-Bench](https://aiwiki.ai/wiki/terminal_bench) 2.0 run using [Claude Opus 4.5](https://aiwiki.ai/wiki/claude_opus_4_5) that he submitted to the leaderboard in December 2025 [2].

Sessions in Pi are trees: a user can rewind to any earlier message, branch from it, and later return, which supports workflows such as running a code review in a branch and bringing only the findings back into the main context [4][16]. The pi-ai layer serializes sessions to JSON and supports handing a conversation from one provider's model to another mid-session [2].

### Deliberate omissions

Much of Pi's design is defined by what it leaves out. It has no support for the [Model Context Protocol](https://aiwiki.ai/wiki/model_context_protocol), which Zechner considers too expensive in context tokens compared with CLI tools the agent reads documentation for on demand; no built-in to-do list; no plan mode; no background process manager (users are pointed at tmux); and no dedicated sub-agent tool [2][4]. It also runs in what Zechner calls "YOLO mode" by default: there are no permission prompts, and the agent has unrestricted access to the filesystem and shell under the invoking user's privileges. He argues that permission systems in other agents are "mostly security theater" once an agent can write and run code [2]. The project README states plainly that Pi has no built-in permission system and documents container and sandbox patterns for users who want isolation [1].

### Extension system

Instead of shipping features, Pi ships an extension mechanism designed so the agent can extend itself. Extensions can register tools and slash commands, render custom TUI components, and persist state into session files; the system supports hot reloading, so the agent can write an extension, reload it, and test it in a loop, guided by documentation that ships with the tool [4]. Ronacher, describing his own setup, noted that extensions had been used for everything from question-answering dialogs and local issue trackers to a demonstration that runs Doom inside the terminal UI [4]. The pi.dev site hosts documentation along with directories of community packages and models [16].

## Adoption

The repository grew to about 85,000 stars within a year of its creation [1]. Besides OpenClaw, third parties have built Emacs and GUI front ends, alternative-language reimplementations, and provider and search extensions on top of Pi, several of which have been shown on Hacker News [18].

[MiniMax](https://aiwiki.ai/wiki/minimax) adopted Pi for its coding product. The company's June 1, 2026 [MiniMax M3](https://aiwiki.ai/wiki/minimax_m3) release post said [MiniMax Code](https://aiwiki.ai/wiki/minimax_code) "is built on a harness based on the outstanding open-source community projects [OpenCode](https://aiwiki.ai/wiki/opencode) and Pi" and that MiniMax planned to open-source the result [13]. A rebuilt core labeled "MiniMax Code 2.0" began rolling out in the desktop app on July 13, 2026, with the changelog citing p95/p99 first-token latency cuts of over 90 percent [15]. On August 7, 2026, MiniMax announced the MiniMax Code 2.0 launch, describing it as "rebuilt on the open-source @pidotdev Pi Agent framework" [14].

[Databricks](https://aiwiki.ai/wiki/databricks) included Pi in a July 8, 2026 study that benchmarked coding agents on its multi-million-line internal codebase. The authors reported that "the harness a model is called from dramatically impacts cost and quality" and that "in many cases, simple harnesses like Pi performed best on our workloads"; running the same model with the same thinking effort through Claude Code or [Codex](https://aiwiki.ai/wiki/openai_codex) versus Pi changed cost per task by more than 2x in some cases while quality stayed the same, with Pi sending about three times less context per turn [11]. Earendil's own commentary on the study said that, combined with [Claude Opus 4.8](https://aiwiki.ai/wiki/claude_opus_4_8) at the xhigh thinking setting, Pi had the highest overall pass rate in the Databricks figures at a lower cost than Claude Code and Codex [10]. That Earendil post reached the Hacker News front page with over 500 points on August 4, 2026 [18].

[Shopify](https://aiwiki.ai/wiki/shopify) engineer David Cortés described building pi-autoresearch, an autonomous optimization loop, as a Pi extension by asking Pi to write it against its own extension documentation; Shopify reported internal results including unit tests running 300 times faster and React component mounting 20 percent faster [10][12].

Maintenance of the project has had to adapt to the same agent wave it rides on. New issues and pull requests from first-time contributors are auto-closed by default and reviewed by maintainers daily, a policy Zechner ties to the influx of low-quality agent-generated submissions [1][6]. Ronacher wrote in May 2026 that issues which are "5% human and 95% clanker-generated" and confidently wrong had become the most frustrating failure mode, because Pi treats a wrong diagnosis in an issue as evidence rather than rumor [9].

## See also

- [Coding agent](https://aiwiki.ai/wiki/coding_agent)
- [Agent harness](https://aiwiki.ai/wiki/agent_harness)
- [OpenClaw](https://aiwiki.ai/wiki/openclaw)
- [opencode](https://aiwiki.ai/wiki/opencode)
- [Claude Code](https://aiwiki.ai/wiki/claude_code)
- [MiniMax Code](https://aiwiki.ai/wiki/minimax_code)
- [Terminal-Bench](https://aiwiki.ai/wiki/terminal_bench)

## References

1. earendil-works/pi repository (README, LICENSE, and repository metadata: created August 9, 2025; MIT; TypeScript; 85,250 stars and 10,574 forks as of August 8, 2026). GitHub. https://github.com/earendil-works/pi
2. Mario Zechner. "What I learned building an opinionated and minimal coding agent". mariozechner.at, November 30, 2025. https://mariozechner.at/posts/2025-11-30-pi-coding-agent/
3. @mariozechner/pi-coding-agent package metadata (first published November 12, 2025). npm registry. https://www.npmjs.com/package/@mariozechner/pi-coding-agent
4. Armin Ronacher. "Pi: The Minimal Agent Within OpenClaw". lucumr.pocoo.org, January 31, 2026. https://lucumr.pocoo.org/2026/1/31/pi/
5. Earendil Inc. "Announcing Pi & Lefos". earendil.com, April 8, 2026. https://earendil.com/posts/announcing-pi-and-lefos/
6. Mario Zechner. "I've sold out". mariozechner.at, April 8, 2026. https://mariozechner.at/posts/2026-04-08-ive-sold-out/
7. Earendil Inc. "A Reflection on our Announcement Today". earendil.com, April 8, 2026. https://earendil.com/posts/announcement-reflection/
8. Armin Ronacher. "RFC 0015: Pi Licensing". rfc.earendil.com, March 30, 2026. https://rfc.earendil.com/0015/
9. Armin Ronacher. "Building Pi With Pi". lucumr.pocoo.org, May 24, 2026. https://lucumr.pocoo.org/2026/5/24/pi-oss/
10. Earendil Inc. "Pi, Minimal and Performant". earendil.com, August 4, 2026. https://earendil.com/posts/pi-autoresearch-and-databricks/
11. Vinay Gaba, Ankit Mathur, Rishabh Singh, Patrick Wendell, Matei Zaharia. "Benchmarking Coding Agents on Databricks' Multi-Million Line Codebase". Databricks Blog, July 8, 2026. https://www.databricks.com/blog/benchmarking-coding-agents-databricks-multi-million-line-codebase
12. David Cortés. "Autoresearch isn't just for training models". Shopify Engineering, April 15, 2026. https://shopify.engineering/autoresearch
13. MiniMax. "MiniMax M3: Frontier Coding, 1M Context, Native Multimodality - All in One Model". minimax.io, June 1, 2026. https://www.minimax.io/blog/minimax-m3
14. MiniMax Agent (@MiniMaxAgent). "Today we're launching MiniMax Code 2.0. Rebuilt on the open-source @pidotdev Pi Agent framework...". X, August 7, 2026. https://x.com/MiniMaxAgent/status/2085679716976243045
15. MiniMax. "MiniMax Agent Changelog" (v3.0.48-v3.0.49, July 13, 2026: "MiniMax Code 2.0: All-New Core"). agent.minimax.io. https://agent.minimax.io/docs/changelog
16. Pi project website and documentation. pi.dev. https://pi.dev/
17. @earendil-works/pi-coding-agent package metadata (first published under the Earendil scope May 7, 2026; v0.84.1 current as of August 7, 2026). npm registry. https://www.npmjs.com/package/@earendil-works/pi-coding-agent
18. "Pi's Minimalism Is Its Advantage" (540 points, August 4, 2026). Hacker News. https://news.ycombinator.com/item?id=49176038
19. "What I learned building an opinionated and minimal coding agent" (421 points, February 1, 2026). Hacker News. https://news.ycombinator.com/item?id=46844822