Delicious Pi, served locally — a real coding agent, running on models served by your own Ollama.
No API key, no account, and no inference request that leaves your machine. Pi Taster gives the agent a place to work and you a way to watch it: describe what you want in the chat panel, and it writes real source files to disk, runs the dev server, and shows you the result running in a preview panel.
Every file it writes becomes a git commit, so anything it does can be undone — including the changes it makes to its own source.
The workspace is a dock: chats, files, history, a terminal and a live preview, dragged into whatever splits and tabs suit the work, and remembered per app.
You'll need Bun, Ollama, and a model that supports tool calling.
ollama serve
ollama pull qwen3-coder:30b # or llama3.1, gpt-oss, mistral-nemo
bun install
bun run devOpen Settings → General, pick your model, and create your first app. Full setup notes, the command list, and troubleshooting are in Getting Started.
- Sandboxed sub-apps. Each one is a real project in
~/.keylimepi/apps/, with its own git repo. The agent can't reach outside the one that's open. - A live preview. Run the app under the chat, click any element to attach it — markup, styles, screenshot — to your next message.
- Approval you can see. Every tool call is gated, inline in the transcript, with a diff attached before you say yes.
- A history you can rewind. Every write auto-commits. Browse the log, roll back, branch, merge.
- Compiler errors the agent reads. A TypeScript service checks each write and tells the agent what it just broke.
- Skills and MCP sources. Reusable instructions on demand, and third-party tools bridged in.
- A visible context budget. Local models have small windows. The meter shows where yours is going.
The full tour, with screenshots, is in What Key Lime Pi Does.
A Bun workspace: the Electron app in apps/electron/, shared types in
packages/core/, and business logic in packages/shared/. Pi owns the agent loop
and its built-in tools; Key Lime Pi adds the permission gate, path confinement, git
auto-commit, code intelligence, and the MCP bridge. More in
Architecture.
Pi ships no sandbox, so confinement is Key Lime Pi's job and lives in one file:
permission-gate.ts. What it
stops, what it doesn't, and what's left open on purpose is written up in
Security.
| Document | What's in it |
|---|---|
| Getting Started | Install, model setup, commands, troubleshooting |
| What Key Lime Pi Does | The feature tour, panel by panel |
| Architecture | Monorepo layout, the stack, who owns what |
| Security | The confinement boundary, permission modes, known gaps |
| AGENTS.md | The deep version: design decisions and the bugs behind them |
| CLAUDE.md | Claude Code–specific workflow on top of AGENTS.md |
| docs/plans/ | One plan and one notes document per implementation session |
Key Lime Pi was built in numbered sessions, each with a plan and a record of what actually got built. Session 15 is the move from a hand-rolled Anthropic loop to Pi on local models; Session 17 covers the shell design and visual identity.
MIT © Guy Ettinger