Features

Everything
AFH does.

An agent that plans its own work, splits it across helpers and shows you every line as it lands — with the controls to steer it, stop it and rewind it.

Download for Windows Getting started
01

It plans its own work.

Before it touches a file, the agent decides what "done" means and lays out the steps — where you can see them and change them.

◎

Goals

The agent sets a goal from your request and keeps working until it's met — or tells you what's in the way. Set one yourself with /goal. Achieved goals are kept under the Goal button.

☰

A to-do list you can touch

Every step shows above the message box. Add, reorder or tick items yourself — the agent sees your edits on its next turn.

Read the auth code
Add session handlingnow
Write the tests
?

It asks instead of guessing

When a choice is really yours — which library, which design — it pauses and asks you multiple-choice questions, then carries on with your answer.

⌁

Picks up after a restart

Close the app mid-run, or lose power: when AFH starts again, interrupted runs resume by themselves.

❖

It remembers your project

How to run it, how to test it, the gotcha that cost an hour, the way you like things done: the agent saves what it learns to the project's memory, and every new thread starts knowing it. Open Memory beside the Goal button to add, edit or forget anything — or just say “remember…”, or type /remember. Your changes reach the agent on its very next step.

Run the tests with pnpm test
The API needs the dev server on port 4000
Use British spelling in the UInew
◐

Three ways to work

Choose a mode in the message box — or create your own agents with their own prompts, models and tools.

BuildReads, edits and runs code in your project. The default.
PlanInvestigates and proposes a plan without changing anything.
DelegateLeads the work and hands tasks to your delegate models.
02

It splits the work.

A lead agent hands pieces to helpers that run side by side, then checks what they bring back.

⇶

Helpers in parallel

The built-in Helper is a copy of your own model for self-contained side work. Add your own subagents too, and watch any of them from the top bar: what the lead asked, and what came back.

◆

Mix models

Lead with your strongest model and delegate to a fast, cheap one. Each helper can use a different provider.

LeadGPT-6.1 Sol
Helper · testsGLM 5.3
Helper · stylesKimi K3.1
⑂

Git worktrees

Give a thread its own worktree — its own branch and folder — then merge it into your base branch when you're happy. Your main checkout stays clean.

⚡

Fast on many files

Reads, searches and helper tasks run in parallel, and old tool output is trimmed so long runs stay quick and cheap.

03

You watch it work.

Nothing happens behind a spinner. You see the thinking, the tools and the code as they happen.

◉

Live file pane

Tool arguments stream as they're generated, so code appears character by character in the Live pane — not after the fact.

⏱

Every step on a timeline

Thinking time, each tool call, tokens per second and how long each reply took — all in the conversation.

◳

It checks its work in a browser

The browser tool drives a real Chrome or Edge, and screenshots go back to models that can see — so it notices when the page looks wrong.

✎

Errors caught after edits

After the agent edits a file, AFH checks it for errors and hands the results straight back, so broken code is fixed on the spot.

🎙

Images and voice

Paste or drop screenshots and files into the message box. Press the mic and speak — your words are typed in for you (with a Groq or OpenAI speech key).

04

You stay in control.

Decide what it may do alone, stop it at any moment, and undo anything — files included.

⛨

Three approval modes

Ask firstapprove edits & commands
Full accessworks freely
Read onlynever edits or runs

Reading .env files always asks, whatever the mode.

↺

Rewind anything

"Undo to here" rewinds the thread and the files on disk to just before a message, and puts your message back in the box to edit and resend. Or fork the thread from any point.

⚑

Hooks

Run your own commands before or after any tool, or when the agent stops. Exit with code 2 to block a tool call — or to make the agent keep going.

$

Cost you can see

Tokens and cost per thread, budgets that stop a run, and prompt caching on by default for Anthropic models. Subscription plans count as no cost.

05

Built for long sessions.

Jobs that take hours, threads with thousands of messages — AFH is made for them.

∞

Endless threads

When a conversation fills 40, 60 or 80% of the model's context, older parts are summarised in place. You can still read every message; the model reads the summary. Or type /compact.

▶

Background jobs

Dev servers, watchers and long builds run in the background while the agent keeps working. See them with /jobs.

⏲

Schedules

"Every morning, update dependencies and run the tests." Run a prompt every few minutes, hourly, daily, on weekdays or weekly — each run is a new thread you can read. Jobs run while AFH is open.

⇄

Bring your threads

Import threads from OpenCode or Zadloop, and your OpenCode keys, endpoints and MCP servers in one click. Export any thread as JSON or Markdown.

06

Make it yours.

Your colours, your language, your commands — and new abilities the agent can install for itself.

◑

46 themes

29 dark and 17 light, from Sunset and Midnight to Dracula, Nord, Tokyo Night and Catppuccin — or build your own. Switch with a circle, diamond, wipe, fade or blur transition.

ع

English & Arabic

A full Arabic interface with right-to-left layout, switchable live without a restart.

/

Slash commands

Built in: /goal /todo /memory /remember /compact /jobs /new /cost, plus /review, /fix-tests, /explain and /commit. Add your own as Markdown files — .claude/commands works too.

⌘

MCP servers & skills

Plug in local or remote MCP servers, and skills from ~/.afh/skills or your project. Skills made for other agents work as they are.

✦

It extends itself

Ask for a new ability and the agent installs it: skills, MCP servers, slash commands, hooks or custom agents.

07

Build on it.

Everything the app does is a local API. Start an agent from a script, a git hook or CI, and stream what it does.

{ }

Local REST API

Create threads, send messages, read replies and diffs, set goals and approve actions — with one key from Settings → Developer.

Private by design

Yours, on your PC.

AFH runs and stores everything on your computer. There's no account and no telemetry.

  • Threads, settings and undo history live in ~/.afh on your PC.
  • API keys are stored only on this computer.
  • Your code goes only to the model providers you choose.
  • The only other calls: the update check, the model catalogue refresh, and web search or voice if you turn them on.

Ready when
you are.

Free for Windows. Install it, sign in or add a key, and give it something to build.

Download AFH See the models