Shared agent operating substrate for Marcelle Labs.
One source of truth for the agent configuration that would otherwise be copy-pasted across repositories: agent instructions, reusable skills, per-tool config templates and the model-routing policy.
Agent configuration drifts. Each repository ends up with its own slightly different
AGENTS.md, its own tool config, its own idea of which model to route a task to — and
a fix applied in one place never reaches the others.
Symlinking into a shared checkout breaks outside a monorepo, and the Claude Code
settings extends field is not a stable API to build on. So consumers sync instead:
pnpm dlx @marcelle-labs/agent-substrate syncThat renders the canonical templates into the consumer repository as ordinary committed files. The result is reviewable in a pull request, works offline, and does not depend on a shared filesystem layout.
| Path | Purpose |
|---|---|
templates/ |
Canonical per-tool config rendered into consumer repos by sync |
routing/model-routing-table.json |
Cost and success-rate routing policy, with a JSON Schema beside it |
skills/ |
Reusable agent skills published to consumers |
bin/ |
The agent-substrate CLI |
AGENTS.md / CLAUDE.md |
This repository's own agent instructions |
AGENTS.md is the detailed reference: distribution model, what consumers
may and may not override, and how a template change reaches them.
Internal Marcelle Labs tooling, published so consumer repositories can install it. It is
not a general-purpose agent framework, it makes no claim to work outside the tool
configurations under templates/, and it is UNLICENSED — published for use, not
offered for redistribution.