bicced

Sidequest — AI skill for Claude Code

AI community

Cut Claude Code token costs by delegating bulk LLM work to cheap models.

How to install Sidequest

This entry records only its repository, not the path inside it, so there is no exact command to give. Open bicced/sidequest and copy the folder into ~/.claude/skills/, or the file into ~/.claude/agents/.

What Sidequest does

Cut Claude Code token costs by delegating bulk LLM work to cheap models. 600+ models via NanoGPT, OpenRouter, Groq or local Ollama — batched, resumable, with real per-call cost tracking.

Alternatives in AI

Zen MCP Server — The power of Claude Code + Gemini / OpenAI / Grok / OpenRouter / Ollama / Custom Model working as one 11.3k ★
Manifest — Real-time cost observability for OpenClaw agents — track tokens, costs, messages, and model usage 4.1k ★
Axonhub — ⚡️ Open-source AI Gateway — Use any SDK to call 100+ LLMs 3.2k ★

README

sidequest

[![tests](https://github.com/bicced/sidequest/actions/workflows/test.yml/badge.svg)](https://github.com/bicced/sidequest/actions/workflows/test.yml) [![PyPI](https://img.shields.io/pypi/v/sidequest.svg)](https://pypi.org/project/sidequest/) [![Python](https://img.shields.io/pypi/pyversions/sidequest.svg)](https://pypi.org/project/sidequest/) [![License: MIT](https://img.shields.io/badge/license-MIT-blue.svg)](LICENSE)

**Cut Claude Code token costs by sending bulk work to cheap models.** A dependency-free CLI that runs one instruction across thousands of items — classify, extract, summarise, rewrite — and writes the results to a file instead of into your agent's context window.

Works with [NanoGPT](https://nano-gpt.com) (~600 models on one pay-per-prompt key), [OpenRouter](https://openrouter.ai), Groq, DeepSeek, Cerebras, Together, or a local Ollama server. Any OpenAI-compatible endpoint, really.

sidequest map --items listings.jsonl --out verdicts.jsonl \
  --instruction "Classify each listing: apply, maybe or skip. Return verdict and reason." \
  --require verdict,reason --batch 40 --workers 16

1000 items -> nanogpt/qwen3.7-flash  batch=40 workers=16
1000 written, 0 already done, 0 failed
28 calls  $0.0062  0.8 min
-> verdicts.jsonl

That is a real run, not an illustration: **1,000 listings classified for $0.0062 in 49 seconds.** Your agent reads those four lines. The 1,000 results are in a file.

The problem it solves

You have work that is enormous in aggregate and trivial per item: classify 25,000 job listings, extract fields from 4,000 scraped pages, rewrite 900 alt texts, label a dataset. Two expensive mistakes are available.

**Sending it to a frontier model.** The price spread between a capable small model and a frontier one is roughly 100x, and classification does not need frontier reasoning.

**Letting your agent read the results.** This one is quieter and much worse. If Claude Code calls a model and the reply la