@pine_forge/provider-agent

PineForge provider agent CLI — connects your local llama.cpp server to the PineForge coordinator with VRAM-aware model loading.

Install

npm install -g @pine_forge/provider-agent

Or use the one-line web installer (agent + llama.cpp + GGUF download + serve):

PINEFORGE_PROVIDER_KEY='lprov_...' PINEFORGE_MODELS='ornith-1.0:9b' PINEFORGE_COORDINATOR_URL='https://api.pineforge.app' curl -fsSL https://app.pineforge.app/install-provider.sh | bash

Prerequisites

• Node.js 20+
• ~6 GB disk for Ornith 9B Q4 (more for larger models)
• The installer auto-downloads llama-server to ~/.pineforge/bin and GGUF weights to ~/.pineforge/models

Manual override: set PINEFORGE_LLAMACPP_BIN or place GGUF files yourself.

Naming: catalog tag ornith-1.0:9b → file ornith-1.0-9b-Q4_K_M.gguf (auto-pulled from Hugging Face)

Quick start

pineforge-agent setup \
  --key lprov_YOUR_KEY \
  --coordinator https://api.pineforge.app \
  --models ornith-1.0:9b \
  --pull

Commands

• pineforge-agent setup --pull — install llama.cpp, download GGUF, save config, start serving
• pineforge-agent pull — download models only
• pineforge-agent login — save coordinator URL, provider key, models
• pineforge-agent serve — poll for jobs and run inference
• pineforge-agent doctor — check llama.cpp / models and coordinator connectivity
• pineforge-agent status — show config and node stats

Environment

| Variable | Default | Purpose | |----------|---------|---------| | PINEFORGE_MODELS_DIR | ~/.pineforge/models | GGUF directory | | PINEFORGE_LLAMACPP_BIN | llama-server on PATH | Server binary | | PINEFORGE_LLAMACPP_PORT | 8080 | Local API port | | PINEFORGE_CTX_SIZE | auto from VRAM/RAM | Context override | | PINEFORGE_NGL | auto from VRAM/RAM | GPU layer count override (999 = all layers) |

Hybrid loading: when VRAM cannot fit the full GGUF, the agent offloads remaining layers to CPU RAM. Layer count and context are tuned from detected VRAM plus os.totalmem(). Cursor tool jobs target ~200k context when the catalog and hardware allow (capped down if VRAM/RAM is insufficient). Set PINEFORGE_CTX_SIZE or PINEFORGE_NGL to override (legacy: LIBERTYAI_CTX_SIZE, LIBERTYAI_NGL).

libertyai-agent is a legacy alias for the same binary.