Embedded agent runtime
Agent + LM Studio (or any OpenAI-compatible API) run in-process inside the CLI — no separate gateway service.
Capabilities
What the CLI harness does.
Caprigo is a terminal-native operator runtime — embedded agent, structured HUD, HOME playbooks, and real local tools. Built for LM Studio; works with any OpenAI-compatible backend.
Core runtime
Four pillars of the harness.
Multi-pane HUD
Header, Agents, Session, Context, and Input panes mirror how operators actually monitor tool loops and caps.
HOME mission executor
Harness compiles intent into playbooks, bootstraps tools before the first LLM call, and auto-drains when models skip native tools.
Digital body
desktop_* mouse/keyboard/screenshot, WinRT OCR, Playwright browser, shell execution — on your machine, under your permissions.
HUD panes
Each pane has a specific operator job.
Header
Backend + model status at a glance
LM STUDIO ONLINE badge, loaded model, busy pinwheel during loads and tool turns.
Session
Live tool cards and replies
Box-drawn tool cards (╭─ desktop_focus), mission verified lines, structured assistant output.
Context
Caps and mission state
LMS, Browser, Desktop, OCR, Web, Mission kind/playbook, Dialect profile, Brain sticky count, Goal.
Input
Operator control surface
Double-Enter STEER mid-turn injection, /loop /brain /bug /clear /models slash commands.
Also included
Learning, recovery, and operator tooling.
Harness features
- Brain lessons + stumble recovery with sticky tags
- Model Grapple dialect profiles (OpenAI tools vs XML vs action cards)
- Lean prompts + mission-scoped tool catalog for 8GB GPUs
- Bug reporter `/bug` packs for agent handoff
- Session autosave under ~/.caprigo
- Optional Vibes marketplace skill imports
Future: web workspace
Fleet/board orchestration packages remain in the repo for a possible later release. The shipped product is the CLI HUD — npm run build:legacy when revisiting the web stack.