Sovereign agent nodes first. Framework details second.
Plaited is a framework for building local-first sovereign agents.
The current direction is:
- one agent per workspace
- a minimal behavioral core in
src/agent - module-composed behavior for planning, memory, skills, MCP, A2A, and verification
- portable skills under
skills/and workspace installs under.agents/skills - a bootstrap CLI that turns infrastructure decisions into an executable setup surface
Plaited's shipped runtime is centered on:
The core owns:
- behavioral engine setup
- signal installation
- heartbeat and snapshot surfaces
- built-in file and inference handlers
- module installation
Higher-level orchestration belongs in modules.
- AGENT-LOOP.md
- INFRASTRUCTURE.md
- default-modules/program.md
- agent-bootstrap/program.md
- three-axis-modules/program.md
- agent-harness-research/program.md
- src/ — shipped framework code
- skills/ — published and reference skills
- docs/ — concise design docs
- dev-research/ — active program specs
Notable skill surfaces:
Requirements:
- Bun
>= v1.2.9 git
Useful commands:
# Typecheck
bun --bun tsc --noEmit
# Full test suite
bun test
# Bootstrap a local deployment scaffold
bun run ./bin/plaited.ts bootstrap '{"targetDir":".","name":"my-agent"}'
# CLI schema discovery
bun run ./bin/plaited.ts --schemaSkills are the portable agent-facing extension surface.
- published skills live under skills/
- workspace-installed skills should live under
.agents/skills/ - runtime composition still happens through modules
Module-program fanout currently runs through plaited program-runner:
plaited program-runner run '{"programPath":"dev-research/skill-modules/program.md","attempts":3,"parallel":2}'