Infrastructure · Models · live from the API

Every model, every lane, what actually answers.

Read straight off /api/admin/models/map: the keys the server holds, the routing settings, the tunnels it can see, and three usage ledgers.

read …
Wiring

Agents → routing lanes → models

Hover any node to trace its path. Solid gold is answering now. Dashed bone is wired but silent. Dotted gold is the fallback that fires when the lane above it refuses. Node and key states on this diagram come from the live probe, not from a snapshot.

livefallbackwired · silentlegacy id · aliaspicker / opt-in
Settings

Lane settings the server is running with

Environment and vault-backed values that pick a model without a deploy. Keys show presence only.

Usage

What actually ran

Three ledgers. cost_log meters server-side API calls. agent_runs meters Jarvy agent-lane turns, which is the only place the Claude Code operator on Gryff appears. cc_chats counts Coder workspace chats. CLI harnesses, opencode, ElevenLabs and image engines are never metered.

Jarvy turns by engineagent_runs · all time

API calls vs dollars

Lifespansfirst → last metered call or turn

API calls by modelcost_log

API spend by modelUSD · estimate, never a bill

Weekly API callslast 9 weeks

Coder workspace chatscc_chats

Price

The cost ladder

List price per million tokens from the server's own table in setJarvyModel(). Each segment runs from input price to output price, on a log axis.

Not on the ladder: free OpenRouter models, the Ollama box (electricity only), and anything on the Claude Max subscription, which bills nothing per token.

Roster

Every model id the platform knows

Grouped by the key or box that answers for it. Greyed ids are legacy names the DeepSeek API still answers to as aliases of deepseek-flash.

Read this before changing a model
Jesse's agent lane is the Max lane. isJesse() (phone + Telegram id, never role) sends his agent-mode turns to the operator on Gryff, a Claude Code session on the Max plan, engine claude-code:opus. On failure the turn drops to deepseek-fallback, visibly, with one Telegram text per hour. Matt and Marcos stay on DeepSeek by Anthropic's terms.
Claude API call sites are nominal. About 20 places hard-code claude-sonnet-4-6 or claude-haiku-4-5, but jarvyComplete() rewrites every lane except local-dell, claude-smart and anthropic-strict to deepseek-free before it dials out.
Two Anthropic lanes stay wired on purpose. claude-smart (Momentum desk) tries the API key first and falls to DeepSeek V4 Pro. anthropic-strict (financial PII) fails closed and returns nothing rather than reach a third party. Both come alive the moment the API account has credit, with no deploy.
Thanos is two engines. The Coder workspace spawns the droplet's claude CLI. The Master Builder panel and pipeline run OpenClaw agent thanos on DeepSeek V4 Flash via OpenRouter.
Legacy DeepSeek ids are aliases. deepseek-v4-flash and deepseek-chat return 200, served as deepseek-flash. LeadLeopard still names the old id and works. telegram-hulk.js is dead code nothing loads.
Ollama has no fallback. The local-dell lane returns an empty string when the tunnel on :11434 is down, by design. The headless coder on Gryff pulls jobs over HTTPS and runs claude -p with no shell; the droplet re-verifies every result.