Agents ship the code. HQFlow shows you the system.

Agents ship faster than you can reread the repo. HQFlow maps what they built, entry point to failure path.

$npx hqflow init
$npx hqflow open

See HQFlow explain itself.

This real canvas maps HQFlow's own workflow, from the first CLI command to a live architecture review and export.

Try it Click “Agent Maps Code” to inspect its real sources, inputs, outputs, and agent boundary.
Real production canvas 7 stages · 1 outcome · local data only

Move around freely. Nothing in this example leaves your browser.

Reading isn't enough.

In an AI-native codebase the scarce resource isn't generation. It's keeping the architecture in view after every feature ships.

Read the code

Agents ship features faster than you can re-absorb the architecture.

Read the docs

Walls of markdown and HTML. Accurate maybe, but no shape.

See the system

A sourced map of real workflows: entry, decisions, side effects, outputs.

One loop, three moves.

Your agent writes. HQFlow watches and draws. You read the result and ask for the next workflow.

Scaffold

hqflow init drops a .codehq/ into your repo: a strict JSON contract and a SKILL.md your agent already knows how to read.

.codehq/
├── project.json
├── SKILL.md          agent contract
└── workflows/

Author

Your agent traces real code paths and writes workflow JSON: steps, sources, edge cases, tests.

{
  "id": "create-order",
  "category": "logic",
  "sources": [{ "file": "app/api/checkout/route.ts",
               "symbol": "POST" }]
}

Observe

HQFlow watches workflow files and re-renders the canvas live as your agent maps more of the system.

workflow changed
  checkout.json
  6 steps · 7 connections
  canvas updated

An instrument, not a chatbot.

Local-first

Renders from files on disk. No accounts, no telemetry, no runtime network calls beyond localhost. Your code never leaves the machine.

No LLM inside

HQFlow contains no model and no prompts. Your agent is the brain. HQFlow is the instrument that keeps it honest.

Evidence attached

Steps carry their real source files, symbols, inputs, outputs, edge cases and tests. The map stays connected to the code it explains.

Built for review

Open details, trace paths, rearrange the canvas and export the result. HQFlow turns an agent's map into something people can inspect together.

Point it at a repository.

$npx hqflow initscaffold .codehq/
$npx hqflow openstart the local canvas

then ask your agent

“Map how checkout actually works in this repo: steps, sources, failure paths. I’ll review the architecture on the canvas.”

Works with Cursor, Claude Code and Codex. Anything that can read a repo.