brix

A skill runtime for Claude Code. Turn multi-step workflows into reusable slash commands backed by real pipelines — not prompt chains.

Brix combines modular building blocks (Python, HTTP, CLI, MCP) into pipelines with a unified JSON interface. Each pipeline can be exposed as a Claude Code custom slash command — giving Claude powerful, token-efficient skills that run as native processes.

AI coding assistants like Claude Code are powerful — but they're wasteful at multi-step workflows. Consider a real task: "Download the last 50 invoice PDFs from my Outlook."

Without Brix, Claude does this:

1.  MCP: list-mail-messages (filter=Rechnung)          →  wait, context grows
2.  MCP: list-mail-attachments (mail #1)                →  wait, context grows
3.  MCP: list-mail-attachments (mail #2)                →  wait, context grows
    ... 48 more times ...
50. MCP: list-mail-attachments (mail #50)               →  wait, context grows
51. MCP: get-mail-attachment (attachment #1)             →  wait, context grows
    ... for every attachment ...
    Python: save file, generate report

That's ~164 tool calls. Each one sends the entire conversation context back and forth. At ~4,000 tokens per round-trip, that's ~656,000 tokens consumed — for what is essentially a batch job.

With Brix, Claude does this:

brix run download-attachments-broad.yaml \
  -p keywords="Rechnung,Invoice" -p top=200 \
  -p output_dir=/host/root/dev/invoices

One call. 6.7 seconds. 79 PDFs. 17 MB on disk.

We tested two strategies against a real Microsoft 365 mailbox:

| Strategy | What it does | API calls | Time | Result | |----------|-------------|-----------|------|--------| | Without Brix | Claude makes each call individually | ~164 | ~10 min+ | Fragile, context overflow risk | | Targeted | ODa