What it does

10x is a self-hosted coding agent that breaks tasks into multi-step pipelines. Each step auto-routes to a model tier—“superfast” for simple work, “smart” for hard reasoning—so you’re not burning premium tokens on boilerplate. It runs via npm install -g 10x-cli and hooks into your own API keys through OpenRouter.

The interesting bit

The “Superpowers” system is basically Makefiles for LLM workflows: you define steps in Markdown frontmatter, pick a model tier per step, and pipe outputs forward with {{previous}}. The built-ins (/review, /debug, /refactor) are sensible; the real bet is that users will write custom ones in .10x/superpowers/.

Key highlights

• BYOK via OpenRouter; no subscription wall
• Three hardcoded tiers: GPT OSS 20B (20×), Kimi K2 1T (4×), Claude Opus 4 (1×)
• Custom skills and project context via 10X.md and .10x/skills/
• Resume sessions by name; one-shot mode with -x
• MIT licensed, 1,351 stars

Caveats

• The “20× faster” claim is relative to Claude Opus 4 alone; actual wall-clock depends on network, caching, and whether your task actually fits the cheap tiers
• No evidence in the README of how routing decisions are made—“auto-picks” is stated, not explained
• The comparison table frames competitors as “single model,” which undersells Cursor’s tab-completion and Copilot’s inline suggestions; 10x is more agent than autocomplete

Verdict

Worth a spin if you’re already juggling OpenRouter keys and want structured, repeatable AI workflows without vendor lock-in. Skip it if you want deep IDE integration or don’t want to manage another YAML-ish config layer.