The open-source AI IDE you actually own
A local-first harness for coding agents, built on Zed.
Bring your own model keys — or run the models on your own machine. No telemetry, no hosted plan.
Download ·
Documentation ·
paddleboard.dev ·
Discussions
PaddleBoard keeps Zed's speed, GPU-accelerated rendering, LSP, debugger, terminal, and git workflow, then layers on the pieces a modern coding agent actually needs: a persona system that lets you define who your agent should be, one-command serverless deploys, sandboxed code execution, sandboxed MCP servers, an embedded browser, step-by-step tool approval, and a live view of every agent thread in the workspace.
Status: beta (0.3.x). Signed, notarized macOS builds (Apple Silicon), a Linux x86_64 tarball and an unsigned Windows x86_64 preview zip are available on the Releases page — or build from source (see below).
✅ Signed & notarized. macOS downloads are code-signed with an Apple Developer ID and notarized by Apple, so they open without a Gatekeeper warning. Linux ships as an x86_64 tarball. Windows ships as an unsigned preview zip: SmartScreen warns on first run, there is no installer or in-app update, and managed Local Models are unavailable — see docs.paddleboard.dev/windows.
What PaddleBoard does differently
Everything you rely on still works — and these additions are all reachable from the command palette or panel bar:
-
Personas — tell the agent who to be. Drop a PERSONA.md at your project root describing who the agent should mimic — a Senior Developer, an SRE, a QA Engineer — and every new agent thread adopts that identity automatically: its voice, values, and what it pushes back on. Plain prose works; frontmatter is optional. Keep a library of <name>.persona.md files in .claude/personas/ (project or ~/.claude/personas/ user-wide) and switch per thread with the picker in the agent panel — or just ask mid-conversation ("be my QA tester") and the agent switches itself via its adopt_persona tool. Three starter roles ship in the AI Dock's Personas tab with one-click install. Personas ride the system prompt, so they work identically with every LLM provider. Skills define what your agent can do; a persona defines who it should be.
-
Local Models — run a model with no API key. Turn on Run locally, managed by PaddleBoard in the AI provider settings and pick a model (Gemma 3 4B, ~2.5 GB, or Gemma 3 1B for low-RAM machines). PaddleBoard ships a pinned llama.cpp server in the app, downloads the model once with a progress bar, and runs it on a private 127.0.0.1 port; it then appears in the agent's model picker like any other. No Ollama, no separate install. macOS Apple Silicon and Linux. Prefer your own server? The same provider still connects to a llama.cpp server you run yourself.
-
Semantic search — find code by meaning, locally. Turn on "paddleboard_rag": { "enabled": true } (off by default) and the agent gains a semantic_search tool that finds where a concept lives when you don't know the keyword to grep for. Embeddings come from a built-in EmbeddingGemma model (~0.33 GB, downloaded on first use) on the same local llama.cpp stack, and the index is a small SQLite file that updates only the files that changed. Nothing leaves your machine. Optionally point it at your own Postgres + pgvector for large codebases.
-
Spotter — routine work on a local model, Claude or Gemini only when it's needed. Name your tiers cheapest first (say Gemma on Ollama, then Gemini Flash, then Claude Sonnet) and give the project a check command such as cargo test -p my_crate. At the end of every agent turn, PaddleBoard runs the check in the agent's own terminal, under the same sandbox and permission prompt. A failure goes back to the same model once; a second failure stops the run with a Blocker Summary (the command, its output, the files changed, what was tried) and offers the next tier in one click. The stronger model continues in the same thread with the whole history. Escalation asks by default; moves up on its own, up to a ceiling you set. Off until you enable it: in settings, or per project in . See .
以上为 README 的开头部分。其余内容请阅读原文。 查看完整 README ↗
