MIT-licensed tools that make agent runtimes easier to connect, inspect, secure, and trust.
Companion for Claude · Bobby Browser · Ableton MCP · 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.
| 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 | Clone and install on macOS with Ableton Live and Node.js. |
| MCP Eval | brew install cavi-ai/tap/mcpeval |
| CAVI API Client | npm install @cavi-ai/api-client |
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 API Client | Gateway-agnostic TypeScript client — agent runtime infrastructure for typed HTTP, WebSocket, and SSE. |
| CAVI Plugins | Host-neutral plugin catalog across Claude, Codex, Gemini, OpenCode, and AgentSkills. |
| Obsidian Agent | Portable, CLI-powered Obsidian workflows across agent hosts. |
| Antigravity for OpenClaw | OpenClaw provider plugin for Google's Antigravity CLI. |
- 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
