aiwg
Version:
Deployment tool and support utility for AI context. Copies agents, skills, commands, rules, and behaviors into the paths each AI platform reads (Claude Code, Codex, Copilot, Cursor, Warp, OpenClaw, and 6 more) so one source of truth works across 10 platfo
59 lines (52 loc) • 2.33 kB
YAML
# Hermes Agent + AIWG — Minimal MCP Configuration
#
# This config connects AIWG as an MCP sidecar.
# AIWG exposes exactly 5 tools: workflow-run, artifact-read, artifact-write,
# template-render, agent-list. Schema overhead: ~3,000 tokens.
#
# Architecture: Hermes (host) → MCP → AIWG (sidecar)
# Hermes owns: conversation, memory, sessions
# AIWG owns: .aiwg/ artifacts, workflows, templates
#
# Usage:
# 1. Copy the mcp_servers and delegation blocks into ~/.hermes/config.yaml
# 2. Config reloads automatically (v0.4.0+) — no restart needed
# 3. Verify: ask Hermes "what AIWG tools are available?"
mcp_servers:
aiwg:
command: "aiwg"
args: ["mcp", "serve"]
# Delegation config — route AIWG workflows to a coding-optimized model
# while the parent stays on hermes3 for conversation.
delegation:
model: "ollama/qwen2.5-coder:14b" # Best tool call accuracy for structured output
max_iterations: 50
# Recommended model strategy — two roles, two models:
#
# Parent (conversation, soul, memory):
# hermes3 — best for roleplay/soul features (recommended)
# mistral:7b — solid general-purpose alternative
# llama3.2:3b — lightweight, fast on low VRAM
#
# Coding / tool calls (via delegation.model config):
# qwen2.5-coder:14b — best tool call accuracy, recommended for AIWG workflows
# qwen2.5-coder:7b — smaller variant, still excellent tool calls, lower VRAM
# qwen3.5:9b — vision + 256K context, strong structured output (8GB VRAM)
# qwen3:8b — strong structured output, thinking/non-thinking modes
# phi4-mini — compact alternative for low VRAM
#
# Note: Qwen models have the best tool call accuracy open-weight. Use them
# for any AIWG workflow involving structured output or function calling.
#
# Pull recommended pair: ollama pull hermes3 && ollama pull qwen2.5-coder:14b
# Compression settings for local hardware (12GB VRAM)
compression:
enabled: true
threshold: 0.30 # Lower than default 0.50 — avoids abrupt context cliffs
summary_model: "ollama/qwen2.5-coder:7b"
summary_provider: "custom"
summary_base_url: "http://localhost:11434/v1"
# Context window (auto-detected from /models endpoint, or set manually)
# model:
# context_length: 32768
max_tool_rounds: 10