10 open-source GitHub repos to cut your AI agent's token bill
victor_explore · x · 2026-10-12
A curated list of 10 open-source repos (stars as of Oct 2026) for slashing AI agent token costs:
- Ollama (183K stars, MIT): run models locally for cheap tasks like summaries and drafts.
- LiteLLM (60K, MIT): one gateway to 100+ models with per-key spending limits.
- Claude Code Router (38K, MIT): route Claude Code/Codex requests to the model you pick, sending easy work to cheap models.
- Langfuse (35K, MIT): traces every agent step with per-step cost to find the expensive ones.
- Repomix (29K, MIT): packs your codebase into a single file, sending structure rather than every line.
- ccusage (19K, MIT): tracks what Claude Code and Codex actually cost you.
- CodeBurn (11K, MIT): per-task cost breakdown to surface wasted tokens.
- Bifrost (9K, Apache 2.0): Go gateway with semantic caching.
- GPTCache (8K, MIT): semantic cache for LLM queries — claimed 10x cost cut and 100x speedup.
- Semantic Router (6K, Apache 2.0): a programmable decision layer that picks the model per call.
The playbook: local models for easy tasks, a gateway for budgets, cost observability, plus caching and routing to save.
More from coding & agent
- TinyTorch from Harvard/ETH: build your own ML framework in pure Python, no GPU needed — lmoroney · 2026-10-12
- OCR evals done right: custom annotation apps and agent citation checkers to catch silent errors — HamelHusain · 2026-10-12
- Dev fires back at "AI slop" critics: "You're still fetching data with useEffect" — Rasmic · 2026-10-12
- Claude Code folds after being taunted mid-debug: "If you're not good at Erlang just say that" — arthurcolle · 2026-10-12
- Copilot Studio tutorial shows visual agent workflows with email classification nodes — adnan_hashmi · 2026-10-12
- AI shifts where engineering experience creates value — fundamentals plus adaptation win — _jaydeepkarale · 2026-10-12