Skip to content

YunusPi: a coding agent harness with persistent project memory and bounded subagents

Open-source software · 2 min read · updated

YunusPi is an independently maintained coding-agent harness with its own core (historically derived from Pi 0.85.1). You bring your own model provider and credentials; the harness supplies the rest: a large tool and skill collection that stays out of context until discovered, persistent project memory, bounded subagents for parallel work, an advisory layer, safety hooks and unified cost accounting. One agent stays responsible for the outcome.

Platform
Linux, Windows via WSL2 Ubuntu, macOS via an Ubuntu VM
Language
JavaScript
Licence
MIT
Requires
  • Node.js 22.19+ and npm
  • git, curl, bash, ripgrep, python3, build-essential, bubblewrap
Run
See the README for the install steps
View on GitHub ↗ Release notes v0.18.0 ↗
YunusPi project intelligence viewer: a dependency graph around a Checkout component with filters and evidence panels.

How does a YunusPi session work?

Start yunuspi in your project directory, choose an available model and describe the work. You do not select a workflow first.

  • Start small. Main sessions expose core editing, inspection, coordination and quality tools. Specialised tool schemas and the full skill catalogue stay out of the initial model context.
  • Discover when useful. The agent can browse capability groups, search short descriptions and load selected tools or read a relevant skill.
  • Get gentle reminders. The first prompt carries a brief, once-per-session invitation to look over useful capabilities. /reminder <text> sets a recurring instruction that is repeated every five minutes at the next active turn boundary; /reminder list shows them and /reminder clear stops them.
  • Keep responsibility clear. Safety hooks enforce access and mutation boundaries; quality checks track evidence. A suggestion, a tool call or agreement between subagents is not proof that the work is correct.

What does discovery look like in practice?

These are agent tool calls, not terminal commands:

// Search short previews without loading full tool schemas.
tool_search({ query: "browser screenshots" })

// Enable a tool after choosing it. This does not execute it.
tool_search({ names: ["browser_session"] })

// Find a workflow without reading the entire skill collection.
skill_review({ action: "search", query: "voxel scene" })

Results are paginated, with three matches by default.

Which platforms and versions does it support?

Linux directly, Windows through WSL2 Ubuntu and macOS through an Ubuntu virtual machine; see the requirements above. Installation builds repository-owned source, and updates accept only explicitly selected YunusPi source. yunuspi --core-info shows the active identity. Version 0.18 added a smaller direct-task tool set, compact assurance receipts with native test diagnostics, optional bounded JEV writing and design triage, more precise SEO checks, decoded motion and local glTF/GLB preflight, and explicit-target network and Linux diagnosis; the repository has a verification report with measured results and limits.

Quick answers

Which model does YunusPi use?
You bring your own model provider and credentials. The harness supplies the tools, memory, subagents, advisory layer and safety hooks around the model.
Is YunusPi a fork of Pi?
Its README calls it an independent lineage historically derived from Pi 0.85.1; YunusPi Core starts its own version lineage at 0.1.0. yunuspi --core-info shows the active identity.
Does it run on Windows or macOS?
Windows through WSL2 Ubuntu and macOS through an Ubuntu virtual machine, according to the project requirements. Linux is the direct target.

More on Downloads