Ollama Launches Pi: The Minimal Coding Agent That Powers OpenClaw Is Now Yours to Customize

Key Takeaways

• Ollax`ma’s new ollama launch pi command lets any developer spin up Pi — a minimal, fully customizable AI coding agent — in a single terminal line. At launch, Pi requires no configuration. 
• Built on roughly 4,000 lines of TypeScript, the system follows a modular monorepo structure featuring packages such as pi-ai and pi-agent-core. These are the same primitives that power the popular OpenClaw agent framework.​
• The default cloud model is Kimi K2.5, a 1-trillion-parameter Mixture-of-Experts model with a 256K context window. Users can access Pi through the command ollama run kimi-k2.5:cloud, priced at just $0.60 per million input tokens and $3.00 per million output tokens. Consequently, it is roughly 9× cheaper than Claude Opus 4.5.
• Pi supports more than 2,000 models from leading providers—including Anthropic, OpenAI, Google, Ollama, and Groq—and lets developers extend its capabilities with custom extensions, skills, prompt templates, themes, and self-written plugins

Quick Recap

Ollama officially announced the launch of ollama launch pi, making Pi — a minimal AI coding agent — directly accessible from the Ollama CLI with zero configuration. Mario Zechner developed Pi, the TypeScript-based toolkit that powers the OpenClaw agent ecosystem. This gives developers a production-grade starting point to build their own personalized coding agents. Ollama announced the update on February 25, 2026, through its official X (Twitter) account and LinkedIn page.

Under the Hood: What Makes Pi a Developer’s Dream Toolkit?

Pi is not just another coding agent wrapper — it is a composable, layered toolkit designed for developers who want control. At its core, the pi-ai package handles unified LLM communication across providers, while pi-agent-core adds tool-calling and agentic loop management. Developers can install Pi globally via npm using the command npm install -g @mariozechner/pi-coding-agent, and configure it easily through a straightforward ~/.pi/agent/ directory structure.

The agent ships with powerful out-of-the-box primitives: tool definitions, session persistence, multi-provider model routing, and a terminal UI layer. From there, developers can install packages like @ollama/pi-web-search (web search and fetch tools) or ask Pi to write its own extensions autonomously. Ollama founder Jeffrey Morgan highlighted Pi’s elegant minimalism and strong performance on small parameter models like the Qwen 3.5 series.

When paired with kimi-k2.5:cloud, Pi inherits a 1-trillion parameter MoE brain built on ~15 trillion mixed visual and text tokens. It also gains Agent Swarm capabilities that decompose complex tasks into parallel sub-agent execution paths. This means a developer can, for example, issue a single natural language instruction. Then Pi will spawn multiple sub-agents working simultaneously on different parts of a coding problem.

The Rise of the Self-Hosted Coding Agent Era

2025 was widely described in the developer community as the year LLMs “became useful,” and early 2026 has marked the transition to coding agents becoming the dominant paradigm. The timing of Pi’s launch is no coincidence — it arrives as OpenClaw, Aider, Claude Code, Codex CLI, and Kiro CLI are all competing fiercely for developer mindshare in the local-first, privacy-preserving agent space.

Ollama’s ollama launch command family (which debuted January 23, 2026) already supports Claude Code, Codex, and OpenClaw as one-click launchable agents. Adding Pi deepens this ecosystem into a bring-your-own-agent layer — a lower-level primitive that lets developers escape vendor lock-in entirely. With OpenClaw requiring a minimum 64K context window and Pi running comfortably even on small models like Qwen3:1.7b, the two tools occupy complementary rungs of the complexity ladder. The rising demand for fully local, zero-API-cost setups — especially on devices like the Raspberry Pi — validates this modular, extensible approach.

Competitive Landscape: Pi + Kimi K2.5 vs. the Field

Pi wins decisively on developer composability — it is the only tool explicitly designed to be a reusable foundation for building entirely custom agents rather than a fixed product. Aider, however, remains the safest and most predictable option for teams that need granular context window control and change-approval workflows. For users who want the most autonomous, “just-go-do-it” agentic power out of the box, OpenClaw remains the benchmark — but Pi gives you the Lego bricks to surpass it on your own terms.

Sci-Tech Today’s Takeaway

In my experience covering the developer tooling space, the most durable products aren’t the ones with the most features at launch — they’re the ones that function as platforms. And that’s exactly what Pi feels like to me.

I think this is a big deal because Ollama has essentially created a zero-friction on-ramp to agent engineering. One command — ollama launch pi — and you’re running a coding agent backed by a 1-trillion-parameter multimodal brain. Moreover, it is at a price point that’s 9× cheaper than the incumbent cloud offerings. That’s not incremental; that’s a category shift.

What I genuinely find exciting here is the self-extension mechanic. The fact that you can ask Pi to write extensions for itself is a quiet but profound capability. I generally prefer tools that grow with me rather than constraining me to a fixed feature set. Pi’s TypeScript monorepo architecture, which already powers OpenClaw, gives it a credibility floor that most new agent frameworks don’t have on day one.

Is this bullish for the local-AI ecosystem? Absolutely. Pi doesn’t compete with OpenClaw — it underlies it, and now it’s in the hands of every developer with Ollama installed. My verdict: for indie developers, researchers, and teams in privacy-sensitive environments, ollama launch pi –model kimi-k2.5:cloud may well become the new default starting point for agent development in 2026. Watch this space closely.