quatcode agent · v0.2.0Now on npm

The terminalhas an agent.

quatcode agent reads your project, plans the work, runs real tools with your approval and verifies its own output — every step visible, right in your terminal.

View on npm

CLIOpen sourceMIT

quatcode agent — live runrunning

timeline

read_filesrc/components/Hero.tsx0.2s

search“premium” · 12 matches0.1s

edit_fileHero.tsx · +42 −180.4s

run_commandnpm run build2.1s

screenshotlocalhost:3000 · captured1.1s

verifytypecheck · build · visual diffdone

$ quatcode "make the hero feel premium"

→ project: next.js 16 · 178 files · pnpm

→ plan: motion + split text + sound

→ editing 3 files · +126 −54

⚠ permission: run_command — allow? y

✓ verify: typecheck · build · diff clean

done — 4 steps · 12.4k tokens

02— Capabilities

Everything it needs.

A small engine with sharp tools. Nothing decorative — only what shipping software actually requires.

01brain

Project brain

Scans the stack, scripts, dependencies and key files — and ranks what matters for the current task.

agent01 / 12

02modes

Plan & build

Tab switches between a read-only plan mode and a full build mode. The header always shows where you are.

agent02 / 12

03tools

Real tools

Files, shell, git, background dev servers and a headless browser — not just chat.

agent03 / 12

04visual

Visual loop

Starts your app, screenshots the page, looks at the result and fixes what looks wrong.

agent04 / 12

05timeline

Agent timeline

Every step with duration and output. /timeline jumps straight to any message you sent.

agent05 / 12

06sessions

Sessions

Resume recent conversations per project. Rename (ctrl+r), pin (ctrl+p), delete (ctrl+d).

agent06 / 12

07skills

Skills

Reusable instruction files the agent loads on demand — add them from a path, a URL or a paste.

agent07 / 12

08mcp

MCP servers

Connect Model Context Protocol servers by URL or command; their tools appear next to the built-ins.

agent08 / 12

09effort

Reasoning effort

Minimal → high, for models that support it — remembered per model and shown in the header.

agent09 / 12

10safety

Permission-first

Commands and edits ask before they run — allow, always or deny. Plan mode blocks writes entirely.

agent10 / 12

11models

Model-agnostic

Anthropic, OpenAI, Gemini or any OpenAI-compatible endpoint — OpenRouter, Ollama, Groq and friends.

agent11 / 12

12open

Open source

MIT-licensed and published on npm. Read the code, fork it, ship with it.

agent12 / 12

scroll12 capabilities

03— The loop

It sees its own work.

The agent doesn’t stop at “done”. It launches the app, looks at the result, and keeps going until the screen matches the intent.

01
Inspect
Reads the project and builds context before touching anything.
02
Plan
Proposes the steps — in plan mode you approve them first.
03
Execute
Runs tools and edits files with a permission check on the way.
04
Verify
Runs the app, screenshots it, checks itself and iterates.

inspect01

$ quatcode run

→ scanning project

178 files · 3 packages · next.js 16

key files ranked for the task

plan02

$ quatcode run

plan · 4 steps

1 read components/Hero.tsx

2 add motion + split text

3 wire sound + verify build

execute03

$ quatcode run

edit_file Hero.tsx · +42 −18

⚠ permission · run_command

$ npm run build · y

2.1s · ok

verify04

$ quatcode run

start_server localhost:3000

screenshot · captured

✓ typecheck · ✓ build · ✓ diff clean

done — 4 steps

start_server → screenshot → edit → screenshot → verify

03— The loop

It sees its own work.

The agent doesn’t stop at “done”. It launches the app, looks at the result, and keeps going until the screen matches the intent.

Inspect

Reads the project and builds context before touching anything.

inspect01

$ quatcode run

→ scanning project

178 files · 3 packages · next.js 16

key files ranked for the task

Plan

Proposes the steps — in plan mode you approve them first.

plan02

$ quatcode run

plan · 4 steps

1 read components/Hero.tsx

2 add motion + split text

3 wire sound + verify build

Execute

Runs tools and edits files with a permission check on the way.

execute03

$ quatcode run

edit_file Hero.tsx · +42 −18

⚠ permission · run_command

$ npm run build · y

2.1s · ok

Verify

Runs the app, screenshots it, checks itself and iterates.

verify04

$ quatcode run

start_server localhost:3000

screenshot · captured

✓ typecheck · ✓ build · ✓ diff clean

done — 4 steps

04— Stack

Any model. Any tool.

The engine is model-agnostic and the tool surface is open — plug in providers, MCP servers and skills.

Anthropic

Claude

OpenAI

GPT

Google

Gemini

Compatible

OpenRouter · Ollama · Groq

MCP servers

stdio · streamable HTTP · SSE — with automatic fallback.

/mcp add https://mcp.example.com/mcp

Skills

Your own instructions, loaded when the task matches.

/skills add ./review.md

One config

Providers, keys, effort and MCP servers live in a single file.

~/.quatcode/config.json

//05 — Get started

Ship with the agent.

Install it, point it at a project and describe what you want. The agent takes it from there — visibly.

npm install -g @quatcode/agent

MIT · Node ≥ 18 · your keys never leave your machine