model-picker

# Model picker The CLI is `model-picker`. If it isn't on PATH, ask the user to install it with `curl -fsSL https://raw.githubusercontent.com/firecrawl/model-picker/main/install.sh | sh`, or run `bun src/cli.ts` from two directories above this skill's base directory instead (after `bun install` there). It needs `FIRECRAWL_API_KEY` (free at https://firecrawl.dev), from the environment or `~/.config/model-picker/.env`; if a command says a key is missing, ask the user to run `mp setup` in their own terminal (it prompts for keys, so don't run it yourself). Results are cached for 24 hours, so repeat runs cost nothing. ## 1. Turn the task into a score floor Scores are the Artificial Analysis Intelligence Index (0-100; the best models score about 55-58). | Task | `--min-score` | |---|---| | Hard agentic coding, multi-step research, long autonomous runs | 50 | | Everyday coding, writing, analysis, tool use | 42 | | Summaries, extraction, classification, routing | 30 | If the user names the model they use today, pass it with `--vs <slug>`; find slugs with `search <name>`. ## 2. Pick the model ```bash model-picker pick --min-score 42 --vs claude-fable-5-1 model-picker pick --min-score 35 --open # self-hostable only model-picker pick --min-score 42 --max-cost 0.5 ``` Each row is a model nothing else beats on both score and cost, cheapest first. Recommend the cheapest row that clears the floor, and say what the user gives up versus their current model (the `vs` column). ## 3. Pick the provider ```bash model-picker providers <slug> # cheapest first model-picker providers <slug> --sort speed ``` Cost per task includes how many tokens each provider's setup actually burns, so it can rank providers differently from list prices. ## 4. Build with the cheapest model per step When the user wants something built cheaply, rather than just a recommendation: ```bash model-picker build "<task>" --dir <folder> ``` A capable model writes plan.json, then each step runs in OpenCode on the cheapest model good enough for its difficulty, in parallel where steps don't share files. It needs OpenCode on the PATH and `OPENROUTER_API_KEY`, and spends real money on that key, so confirm with the user first. Afterwards, show them `<folder>/REPORT.md`: each step's model, time, tokens and cost, plus a per-model breakdown. Add `--json` to `pick`, `providers` and `search` to read the numbers programmatically.