The framework

Foundry

Foundry wraps Claude Code sessions with persistent memory, project conventions, multi-thread awareness, and correctness checking. It's a nervous system around a capable model — so the agent stops re-learning what you already taught it.

★ Foundry on GitHub Oracle waitlist

The agentic team.

Five roles wrap every message. Defined by where they show up in the lifecycle — not by headcount. One agent or many is an implementation choice, not an architecture one.

the foundry CARTOGRAPHER

CARTOGRAPHER

Routes context

Reads the topology map. Decides what context the message needs. Sees everything, modifies nothing.

Install

git clone https://github.com/inixiative/foundry.git
cd foundry
bun install
bun run setup
bun run start

Requires Bun 1.0+. Pick a provider (Anthropic, OpenAI, Gemini, or Claude Code CLI) during setup.

Design principles

Don't reinvent Claude Code. Wrap it, don't replace it.
Pull over push. Inject predicted context, let the session pull the rest via MCP.
Progressive depth. Basic layers every message, deeper layers only on domain shift or low confidence.
One writer, many readers. The Librarian is the sole writer to thread-state.
Cheap models for decisions, capable models for work. Classifiers and routers run on fast/cheap models.
Every execution is a training step. Memory entries, convention proposals, and signal ingestion happen after each interaction.

License

packages/core — MIT. packages/foundry — BSL 1.1, converts to MIT on 2030-04-06.