🦾

Build. Grade. Improve. Accountably ship.

A tested, swappable-seam engine that turns one sentence into an agent app — and never self-certifies that it's good.

pipx install git+https://github.com/wjlgatech/anyagent.git

0 runtime deps 189 tests Free LLM tiers Human-gated publish Swappable seams

The centerpiece

The loop that never self-certifies

Watch AnyAgent build an app from a sentence, grade it against the goal, improve it, re-grade — and then stop, refusing to ship until a human approves.

anyagent — closed loop

Bounded. Logged. Measured. The grade is mechanical (0–100); the ship decision is human.

Build for $0

Try it live — free, in your browser

Paste a free Groq or Gemini key and watch a real LLM draft an agent spec from your idea — the same free tiers AnyAgent's --provider groq|nim|gemini uses to build for $0.

Get a free key ↗ 🔒 Your key stays in your browser — sent only to the provider, never to us.

One CLI, many jobs

What it does

Every capability is a swappable seam — a real adapter plus a test fake. Add a provider or publisher in one file and one line.

build

anyagent build

Scaffold a working agent app from one sentence — swappable framework target, real tests.

refine

anyagent refine

Build, grade, and improve in a bounded loop until the result clears your target score.

analyze

anyagent analyze

Score any repo's OOP quality 0-100 and name its god-functions and ranked, actionable gaps.

refactor

anyagent refactor

Drive a spaghetti repo toward OOP — tests gate every step, and regressions roll back.

reverse

anyagent reverse

Reverse-engineer software, web, docs, podcasts, video, or ads into a fidelity-scored blueprint.

pain2gain

anyagent pain2gain

Mine real Hacker News + arXiv signals into ranked pains and an evidence-backed product blueprint.

The 10x

Why — it closes the loop

Most tools build an agent app once and walk away. AnyAgent grades the result against your goal, repairs it, and refuses to self-certify.

Build

Scaffold a working app from one sentence, on a swappable framework target.

Grade

A mechanical evaluator scores the result 0–100 against the goal — not vibes.

Improve

The RefineLoop iterates, bounded and logged, climbing toward your target score.

Accountably ship

Publishing is human-gated — --approve is a decision an agent never makes for you.

Built as swappable seams, not a monolith — every piece testable and replaceable. 0 runtime dependencies. 160+ tests.