🔥 Foundry — your local agentic platform

A local-first, Ollama-powered agent + automation library. Runs on your laptop (RTX 4050 6GB / 16GB RAM), no cloud, no API keys. Built for shipping Lantern and automating the boring parts.

foundry/
  setup/        00-install -> 01-models -> 02-verify   (run in order)
  core/         the agent engine (Ollama client, tool registry, ReAct loop)
    foundry/    python package:  python -m foundry "<task>"
    config.yaml central config (model, tools, paths)
  models/       Modelfile for the IQ4 Gemma (tuned for 6GB)
  templates/    ready automations
    dev-autopilot/   commit msgs, PR review, daily digest, crash triage
    pc-automation/   organize downloads, folder watcher
  makers/
    template-maker/  scaffold new templates
    software-maker/  spec -> scaffold -> test -> deploy -> maintain
  data/         vector memory + generated builds (gitignored)

Quick start

# 1. install (venv + deps)
powershell -ExecutionPolicy Bypass -File C:\AI\foundry\setup\00-install.ps1
# 2. models (embeddings now; Gemma GGUF when you download it)
powershell -ExecutionPolicy Bypass -File C:\AI\foundry\setup\01-models.ps1
# 3. verify the whole stack
powershell -ExecutionPolicy Bypass -File C:\AI\foundry\setup\02-verify.ps1

# run the agent on any task
cd C:\AI\foundry\core
..\.venv\Scripts\python -m foundry "summarize the README files under templates"

The brain

• Default: qwen2.5:7b — already installed, fits fully in 6GB VRAM, supports tools. Fast.
• Upgrade: gemma-foundry (REAP Gemma-4-19B @ IQ4_NL) — smarter, runs hybrid GPU+CPU (~10.7GB). Download the GGUF, run 01-models.ps1, then set ollama.model: gemma-foundry in core/config.yaml.
• Avoid for agents: the Qwen3.6-35B @ IQ2_XXS — great chat ceiling, but 2-bit breaks tool-calls/JSON. Keep it as a separate chat model, not the agent brain.

Tools the agent has

read_file, write_file, list_dir (sandboxed to the foundry dir) · run_shell (allow-listed exes) · web_fetch (+ web_crawl if crawl4ai installed) · memory_add / memory_search (local numpy vector store).

Add your own tool

Drop a register(reg, cfg) in core/foundry/tools/, wire it in tools/__init__.py. Each tool is ~10 lines. Crawl4ai/turbovec/qdrant slot in the same way.

Coordinating more tools (crawl4ai, vector DBs, etc.)

• crawl4ai: .\.venv\Scripts\pip install -r core\requirements-extra.txt then python -m playwright install chromium. web_crawl appears automatically.
• Bigger vector store (turbovec/qdrant/chroma): keep the memory_add / memory_search signatures, swap the adapter in core/foundry/tools/vectorstore.py.
• Docker services: add docker to config.tools.shell.allow and let the agent spin up containers (you have Docker installed).

Scheduling

Wrap any template in Windows Task Scheduler, or use the /schedule skill in Claude Code for cloud cron. Example: daily-digest every evening.

schtasks /create /tn "FoundryDailyDigest" /tr "powershell -File C:\AI\foundry\templates\dev-autopilot\daily-digest.ps1" /sc daily /st 19:00