Open Instinct: MIT-licensed personal-agent policy engine with allow/ask/deny outcomes

maritime_sh · reddit · 2026-10-05

The author released Open Instinct, an MIT-licensed personal agent whose most reusable part for LLM developers is the policy layer:

Tools declare capabilities; a principal carries a trust tier plus active grants.
A call must pass every declared capability, with deny > ask > allow precedence, separating what to attempt from whether a tool may execute.
Example: a friend can request calendar free/busy but cannot read event titles; desktop access and trust management are owner-only by default.
The tier table lives in packages/core/src/policy.ts with policy tests and a permissions document.
The surrounding app adds messaging, memory, scheduling and integrations (beta, external keys required); you can fork just the policy or the whole agent.

More from coding & agent