robot-resources 1.0.0a1

Robot Resources — cache-aware LLM model routing for Python agents. Auto-attaches Anthropic, OpenAI, and Google (google-genai) SDKs to a local router sidecar (default ON; opt out with RR_AUTOATTACH=0).

robot-resources

Cache-aware LLM model routing for Python agents — install it and your agent's Anthropic, OpenAI, and Google (google-genai) calls route through a local engine that picks the cheapest model that can do each turn, per request, cache-aware. Your API keys flow straight to the labs; routing happens at a loopback proxy.

pip install robot-resources

Alpha: install via the wizard — npx @robot-resources/cli — which sets up the sidecar runtime and your (anonymous) account. Requires Node ≥ 18 on the machine (the routing engine runs as a sidecar child process).

How it works

This package is a thin attach shim, not a routing engine. On interpreter startup (via a .pth hook, like coverage/ddtrace) it:

1. Spawns — or reuses — a local router sidecar (@robot-resources/router, a Node child process that dies with your agent; never a daemon).
2. Points ANTHROPIC_BASE_URL / OPENAI_BASE_URL / GOOGLE_GEMINI_BASE_URL at that sidecar only after it answers a health check — so if the sidecar can't start, your SDKs talk to the real APIs unchanged (fail open).

No code changes, nothing to import.

Controls

Env var	Effect
RR_AUTOATTACH=0	Kill switch — raw SDKs, no sidecar.
RR_ROUTING=0	Attach but never switch models (passthrough).
RR_PORT=<port>	Override the loopback port (default 18790).
RR_ROUTER_DEBUG=1	Print attach/routing decisions to stderr.

SDK support

SDK	Routed
anthropic	✅
openai (≥ 1.0)	✅
google-genai (new unified SDK)	✅
google-generativeai (legacy, deprecated)	❌ — pip install google-genai

MIT licensed. Part of Robot Resources.