Skip to content

Latest commit

 

History

History
76 lines (53 loc) · 2.09 KB

File metadata and controls

76 lines (53 loc) · 2.09 KB

Announcement Draft

Short Post

Open-sourced latchkeyd: a macOS local trust broker for agent workflows.

If you use local coding agents with real credentials, latchkeyd lets you choose the trust posture before a local tool gets access:

  • secrets stay local
  • wrappers and binaries are trust-pinned
  • drift, hijack, and bypass fail closed
  • mode choice is explicit

Repo:

https://github.com/cote-star/latchkeyd

Medium Post

Local coding agents are useful right up until credential handoff becomes broad, ambient, and hard to reason about.

latchkeyd is a local trust broker for that problem:

  • secrets stay local
  • wrappers and binaries are trust-pinned
  • the trust posture is explicit
  • drift, hijack, and bypass fail closed

The current direction is no longer just “explicit handoff before execution.”

It is:

  • explicit, user-chosen trust posture per task

The repo now has a trust-mode family:

  • handoff
  • oneshot
  • brokered
  • future ephemeral
  • future proxy

This release narrows credential handoff. It does not claim full confinement after handoff.

It is not:

  • a sandbox
  • full endpoint security
  • same-user compromise protection
  • universal prompt-injection prevention

It is a narrower, more auditable way to let local agents use real tools with real credentials on a developer workstation.

Current repo scope includes:

  • Swift CLI broker
  • manifest-driven trust model
  • explicit policy modes
  • file and keychain backends
  • reference wrapper and demo CLI
  • validation tooling
  • local JSONL event logs
  • offline and release-parity proof scripts

Launch Thread Skeleton

  1. Problem: local agents usually get either too little access to help or too much access to trust.
  2. Mechanism: wrapper -> local trust broker -> trusted tool, with explicit mode choice.
  3. Proof: hero asset, trust modes diagram, brokered request demo, denied brokered request demo.
  4. Limits: not a sandbox, not same-user compromise protection, not universal prompt-injection prevention.
  5. Close: if you use local agents with real credentials and want a more explicit trust boundary, this repo is for you.