RAGBAZ / grove / Sylvae
What each backend reuses, what it spends, and why automatic routing refuses to guess.
published · Updated 2026-10-05
01 Sylvae
Ollama runs a pulled model on your own machine through an explicit backend and model name. There is no API key and no per-call fee; the costs are hardware, electricity, maintenance and your attention to whether a small model's output is good enough.
A bare model name is prefixed to the litellm ollama/ convention when omitted. Pull the model before the run; a missing model is an unavailable run with a clear next action, not a mysterious failure.
02 Sylvae
The Anthropic backend uses a configured API key; the claudecode backend reuses an existing Claude subscription through the local CLI with settings trimmed to shed unused bootstrap. Both are real metered inference with real latency.
The documented trimming experiment matters here: identical prompts differed by an order of magnitude in tokens and cost depending on setup. Defaults are not neutral, and Sylvae's flags exist to make the expensive parts visible rather than to promise they are gone.
03 Sylvae
Shellout and opencode invoke real agent CLIs with their own authentication, tools and bootstrap overhead. They suit multi-step work and punish trivial prompts with fixed costs. Neither participates in cheap routing, and neither is selected by auto.
The claudecode backend shares an interactive quota, so a run spends the same allowance used for ordinary work. Hitting that limit is reported as unavailable. That distinction protects both the evidence log and the operator's afternoon.
04 Sylvae
With --backend auto, a declared cheap tier routes locally and a declared agent tier routes to a harness; anything undeclared, including a misspelling caught at load time, resolves to frontier. Per-tier overrides exist because the right target genuinely differs per operator.
Over MCP the guards tighten further: delegation defaults to cheap, the subscription-sharing backend is refused, and auto is refused outright, since tier routing could otherwise select a recursion-risk backend. An operator lifts these with flags; a calling model cannot grant them to itself.
A conversation, not a sales funnel
Ask a question, tell us what you need, offer a contribution or simply show your support. RAGBAZ is a software studio working toward kindness, mutual cooperation and the good of all beings.
Prefer email? ragbaz@proton.me