Caprigo

Local-first agent harness in your terminal. Connect LM Studio (or any OpenAI-compatible API), run caprigo, and operate through a multi-pane HUD — embedded runtime, live tool cards, HOME missions, desktop body. No gateway. No browser dashboard.

Site	caprigoai.com
Credit	b_Radford · Vibes-Coded
Stack	Node · TypeScript · LM Studio · MIT

---

The HUD

Default path: start LM Studio → load a tool model → launch the harness.

Pane	What you see
Header	LM STUDIO ONLINE, model name, busy pinwheel
Agents	Active session, /loop /think /models
Session	User turns, ╭─ tool cards, mission verified
Context	Caps — Desktop, OCR, Web, Brain, Mission, Goal
Input	Double-Enter STEER, /bug /brain /clear

Mission examples

Mission	What happens
Desktop	open notepad and type hello → HOME playbook, OCR verify
Web	Search + fetch with visible tool cards in Session
Write	write_file → browser preview of generated HTML

---

Quick start

Windows (fastest)

git clone https://github.com/doteyeso-ops/caprigo.git
cd caprigo
.\setup.ps1 -LaunchHud

Manual

git clone https://github.com/doteyeso-ops/caprigo.git
cd caprigo
npm install
npm run build
cp .env.example .env   # LM Studio defaults inside
.\launch-hud.ps1       # Windows
# caprigo              # same as caprigo tui

Prerequisites: Node 18+, LM Studio Local Server on :1234, a tool-capable model (e.g. qwen2.5-coder-7b-instruct).

caprigo doctor          # verify backend + .env
caprigo connect         # auto-discover LM Studio on LAN
caprigo setup -i        # interactive .env setup

---

Configure LM Studio

.env at repo root:

CAPRIGO_LLM_PROVIDER=openai_compatible
OPENAI_BASE_URL=http://127.0.0.1:1234/v1
DEFAULT_MODEL=qwen2.5-coder-7b-instruct
CAPRIGO_HARNESS_MODE=1

Ollama works too — see .env.example.

---

Why a harness

Models keep getting better; most wrappers are still a chat box with hidden tools. Caprigo is the execution layer:

• HOME — harness compiles intent into playbooks and runs bootstrap steps before the first LLM reply
• Auto-drain — finishes steps when the model ignores native tools[]
• Digital body — desktop_* mouse/keyboard/OCR on Windows, Playwright browser, shell
• Brain + stumble — sticky lessons, failure recovery, model dialect profiles
• STEER — double-Enter mid-turn injection while a tool loop is running
• /bug — handoff packs under ~/.caprigo/bug-reports/

Lean mode for 8GB GPUs: CAPRIGO_LEAN_TOOLS=1 (see .env.example).

---

CLI reference

Command	Purpose
caprigo / caprigo tui	Launch embedded HUD (default)
.\launch-hud.ps1	Windows launcher (builds if needed)
caprigo doctor	Config + LM Studio probe
caprigo connect	Discover LM Studio, write .env
/loop /brain /bug	In-HUD slash commands

---

Docs

• INSTALL_AND_FIRST_RUN.md — first-run walkthrough
• docs/SCREENSHOTS.md — regen assets (npm run screenshots)
• SOCIAL_LAUNCH.md — share copy
• HandOff.md — agent handoff notes

Future (in repo, not default): web workspace / gateway — npm run build:legacy when needed.

---

License

MIT — see LICENSE.