independent research / open sourcestatus: building_in_public

Open tools for agents that work across runtimes, editors, and environments.

Research in public. Ship under MIT. Make systems easier to inspect, adapt, and trust.

Quick install

claude plugin marketplace add cavi-ai/mlx-agent
claude plugin install mlx-agent

Open sourceMIT by default31 agent skillscanonical playbooksMLX Agent v0.1.0stableVerified Jul 2026manifest-backed

/solutions

Start with the outcome

Find the CAVI tools, skills, and installation path for the job you want to accomplish.

01 / local-models

macOS / Apple Silicon

Run models locally

Discover, verify, and wire local MLX-optimized models on Apple Silicon.

solutionMLX Agentmlx-scout

view_solution()

/work

Current work

Public repositories, practical experiments, and developer tools. Every project is open to inspect, fork, and improve.

01 / TypeScript packagev0.11.0 stable

@cavi-ai/api-client

One typed client. Every agent runtime.

A follower package for compatible HTTP, WebSocket, and SSE runtime surfaces. Typed boundaries make upstream behavior easier to integrate and verify without claiming protocol ownership.

read_docs()view source

verified against npm v0.11.0

runtime-registry.tsstable

import {
  createRuntimeClient,
  createRuntimeProviderRegistry,
} from "@cavi-ai/api-client";

const registry = createRuntimeProviderRegistry({
  modules: [createCodexProviderModule({ apiKey })],
});

const client = createRuntimeClient("codex", {
  registry,
  clientOptions: { baseUrl: "https://api.openai.com" },
});

02plugin marketplace

claude-plugins

Add it once, install any CAVI Claude Code plugin — the marketplace for the whole toolkit.

view_source()

03Obsidian plugin

companion-for-claude

The Obsidian companion: an in-vault interface for Claude, interactive artifacts, conversation history, and local-model fallback.

MITview_source()

04Claude Code plugin

claude-obsidian

The agent-side half of the Obsidian workflow, providing 30 focused skills for research, writing, vault maintenance, and delivery.

MITview_source()

05Claude Code plugin + skill

mlx-agent

Discover, verify, and wire local MLX-optimized models on Apple Silicon for your coding agent.

MITview_product()

/skills

Agent skills

31 agent skills across 2 providing plugins. Start with the plugin that owns the workflow, then browse its focused playbooks.

2 skill providers

30skills

claude-obsidian

The Claude Code plugin behind the Obsidian research, writing, organization, and delivery playbooks. Pairs with companion-for-claude in the vault.

skillsbrowse_provider()

01skill

mlx-agent

The MLX Agent playbook for discovering and configuring local models on Apple Silicon.

skillsbrowse_provider()

view_all_skills()install via marketplace

/research

What we're testing

Our research starts in working software: explicit assumptions, reproducible behavior, and results anyone can examine.

runtime.interoperate()

Runtime interoperability

Agent tools should travel across gateways and providers without rebuilding their foundations each time.

contracts.verify()

Contracts over conventions

Typed boundaries, owned routes, and executable checks make system behavior inspectable instead of implied.

systems.degrade_gracefully()

Graceful degradation

Useful systems should preserve context and return structured gaps when upstream services are incomplete.

workflows.stay_local()

Local-first workflows

Developer tools should meet people where their knowledge already lives, with local control as the default.

license.txt

Built in the open, by default.

Our public work is released under the MIT license. Read the code, test the assumptions, adapt the tools, and contribute what you learn.

explore all repositories