MIT-licensed tools that make agent runtimes easier to connect, inspect, secure, and trust.
Companion for Claude · Bobby Browser · Ableton MCP · CAVI API Client · Documentation
Companion for Claude brings Claude into your Obsidian vault: vault-aware chat, an agent mode whose edits you review hunk by hunk, sandboxed HTML artifacts, and evidence-backed research. The Obsidian Community Store release for Claude knowledge workflows. Connect with Claude Code, your own Anthropic key, or a local model.
Install from the Obsidian store · View source
Bobby Browser is an alpha browser-automation runtime for agents on Firefox and Chromium. Every call is authenticated, capability-scoped, journaled with evidence, and recoverable after a crash, through MCP, ACP, CDP, and SDKs for TypeScript, Python, and Rust.
Install · Documentation · View source
Ableton MCP gives MCP clients local-first control of Ableton Live through a Remote Script bridge, a stdio MCP server, and a CLI with confirm-to-execute mutations: every change to the Live Set is planned first and runs only with a single-use confirmation.
CAVI API Client lets one TypeScript application offer a choice of agent runtimes. It adapts Claude, Codex, AGY, OpenCode, Hermes, and OpenClaw to shared contracts for runs, statuses, stream events, and token usage, so you write one progress renderer and one failure handler instead of one per backend. Gateway-agnostic agent runtime infrastructure over typed HTTP, WebSocket, and SSE; ESM with no runtime dependencies.
Install · Documentation · View source
| Product | Install |
|---|---|
| Companion for Claude | In Obsidian: Settings → Community plugins → Browse, search "Companion for Claude", or open the store listing. |
| Bobby Browser | brew install cavi-ai/tap/bobby-browser |
| Ableton MCP | npm install -g @cavi-ai/ableton-mcp, then ableton-mcp install |
| Secure Agent | Download the macOS app (macOS 14 or later). |
| MCP Eval | brew install cavi-ai/tap/mcpeval |
| MLX Agent | claude plugin marketplace add cavi-ai/mlx-agent, then claude plugin install mlx-agent@mlx-agent |
| MLX Workbench | Download the macOS app (Apple Silicon). |
| CAVI API Client | npm install @cavi-ai/api-client |
| CAVI Plugins | claude plugin marketplace add cavi-ai/plugins |
| Obsidian Agent | /plugin marketplace add cavi-ai/obsidian-agent, then /plugin install obsidian-agent@obsidian-agent in Claude Code |
| Antigravity for OpenClaw | openclaw plugins install @cavi-ai/antigravity |
| Loop Killer | claude plugin marketplace add cavi-ai/plugins, then /plugin install loop-killer@cavi-ai in Claude Code |
| Blender MCP | uv tool install cavi-blender-mcp, then cavi-blender-mcp runtime start to open Blender with the bridge |
| Substance MCP | uv tool install cavi-substance-mcp for the Painter, Designer and Sampler servers |
Open building blocks for secure, local, and inspectable agent systems.
 |
 |
| Secure Agent | MLX Workbench |
| Project | What it does |
|---|---|
| Secure Agent | Local monitor for AI coding agents on macOS: sessions, processes, file activity, and network destinations. Optional guards can stop supported tool calls; an opt-in proxy inspects routed traffic for secret leaks. |
| MCP Eval | Evidence-driven MCP-server evaluation across discovery cost, schema guessability, error honesty, state recovery, and contention — turning agent friction into actionable findings. |
| Project | What it does |
|---|---|
| MLX Agent | Discovers, verifies, and wires local MLX-optimized models on Apple Silicon — Claude, Codex, Antigravity, OpenCode, and AgentSkills. |
| MLX Workbench | Native macOS app for the MLX Agent model lifecycle: converts, verifies, compares, serves, and fine-tunes local models on Apple Silicon. |
| Project | What it does |
|---|---|
| CAVI Plugins | Host-neutral plugin catalog across Claude, Codex, Agy, OpenCode, OpenClaw, Hermes, and AgentSkills. |
| Obsidian Agent | Portable, CLI-powered Obsidian workflows across agent hosts. |
| Antigravity for OpenClaw | OpenClaw provider plugin for Google's Antigravity CLI. |
| Loop Killer | Claude Code plugin that ends stop-hook re-prompt loops: when replies to a re-prompting hook stop doing tool work, it clears the active goal and says so. |
| Project | What it does |
|---|---|
| Blender MCP | Local-first MCP tools for Blender scene inspection, mesh editing, previews, and game asset export, through a bridge that binds to loopback and requires a session token. |
| Substance MCP | Local-first MCP servers for Adobe Substance 3D Painter, Designer and Sampler: layer editing, baking, texture export, batch jobs that export for Unreal, Unity and Blender, and SHA-256 receipts that chain across all three apps. |
- Agent security — recording what local agents read, run, and connect to, with guards you opt into.
- Runtime interoperability — connecting gateways and providers through typed boundaries.
- Local-first workflows — running models and knowledge tools on hardware you control.
- MCP reliability — measuring discovery cost, schema guessability, error honesty, state recovery, and contention.
Every public CAVI-AI project ships with its source and provenance. Read the code, test the assumptions, adapt the tools, and contribute what you learn. More work is in development — we publish it when there is working software and evidence worth examining.
cavi-ai.xyz · all repositories · plugin catalog
MIT licensed · Source published · Provenance attached
