› cavi-ai/

Inspectable tools for local models, agent runtimes, and Ableton Live.

Independent research and open source, from knowledge workflows to a guarded MCP server for your Live Set. Every release ships with its source, verification date, and commit, so you can check it before you adopt it.

Explore CAVIRead documentation
released artifacts
9
agent skills
31
licensed, source published
MIT
Ableton MCPv0.3.1328 tools, 138 read-onlyDocsRelease (opens in new tab)Source (opens in new tab)

/new

New in the catalog

The newest releases. Each links to its source and its release, and to its documentation where it publishes any.

Released · v1.3.0

Secure Agent

An egress-inspection and secret-leak firewall for local AI agents. It watches the process tree, correlates a sensitive-file read with the socket that follows, and flags the leak before your key walks out the door.

  • In-harness secret guard gates PreToolUse reads of keychains, SSH keys, and cloud credentials.
  • Sliding-window correlation engine ties file activity to outbound sockets outside the vendor allowlist.
  • Rotation advisory scores blast radius and writes copy-paste remediation. It is read-only; nothing rotates itself.
  • Go daemon on macOS Endpoint Security, under 2% CPU.
Download (opens in new tab)Source (opens in new tab)
installmacOS 14+
Download SecureAgent-v1.3.0.dmg → Applications
License
MIT
Runtime
Go daemon + Swift menu bar
Hooks
Claude Code · Cursor · OpenCode

/explore

Everything CAVI ships

Nine released artifacts across editors, browsers, runtimes, and packages. Each one links straight to its documentation, source, and install path.

/solutions

Or start from an outcome

Each solution connects the job you want done to the products and skills that complete it.

01Apple Silicon · local inference

Run models locally

Find MLX models that fit your Apple Silicon host, verify them against an installed runtime, and wire a selected model into an agent workflow.

  • MLX Agent
  • mlx-scout
Open solution
02Obsidian · knowledge workflows

Knowledge workflows in Obsidian

Obsidian Agent skills for research, vault hygiene, writing, and handoff, with Companion for Claude as the in-vault interface.

  • Companion for Claude
  • Obsidian Agent
Open solution
03Browser runtime · audit trail

Browser automation with an audit trail

Give each agent a capability-scoped token, drive the browser through one runtime with shared idempotency and evidence contracts, and restart long workflows from a checkpoint instead of replaying side effects.

  • Bobby Browser
Open solution

/how it’s built

Three commitments under every release

Not features, constraints. They hold across the whole catalog, from the security daemon to the docs you are reading.

local-first

Runs on your machine

Daemons, models, and browsers execute locally. No telemetry home, no cloud dependency to inspect a secret or run a model.

inspectable

Boundaries you can read

Capability scopes, allowlists, and hook contracts are explicit config, not opaque policy. You see exactly what an agent may touch.

provenance-backed

Verified before you adopt

Every artifact carries its source repo, release commit, and verification date. The site ingests immutable release assets, never a checkout.

Featured releaseMLX Agent v0.7.0
License
MIT licensed
Verification
Verified Jul 2026