Local-first orchestration for coding agents

AgentForge

A native-feeling macOS command center for scheduling, monitoring, and chaining Claude Code and OpenAI Codex CLI agents from your own machine.

Download for macOS View source

4 schedule modes Bun backend 127.0.0.1 only Claude + Codex

AgentForge task board

1 running/3 queued

Queue

Review scheduler drift

Check delayed and cron tasks after app sleep and wake.

scheduledcodex

Draft Feishu smoke test

Create a compact channel startup checklist.

feishuqa

Running

Refine trace output

Group stream-json events into readable checkpoints.

renderertrace

Done

Ship skill sweep summary

Summarize promoted patterns and candidate skills.

completedskills

Built for repeatable agent work

Run agents like a local automation pipeline.

AgentForge keeps the human in control while making agent work schedulable, observable, and reusable.

Kanban control

Create tasks, watch live output, retry failures, fork successful runs, and keep scheduled work visible.

Flexible scheduling

Run immediately, after a delay, at a specific date, or on cron. Heartbeats can decide when work is worth creating.

DAG workflows

Chain tasks into dependencies so research, implementation, review, and reporting can run as one workflow.

Skill Library

Detect repeated work patterns, distill them into standard SKILL.md files, and load them into Claude Code and Codex.

Live event stream

Read structured output from Claude Code or Codex CLI without tailing terminals or digging through logs.

Local first

The backend listens on loopback only. Your code stays on your Mac; channels and agents connect through local control.

Compounding Skill Library

Your agents get better at your work.

AgentForge can scan completed runs for recurring recipes and pitfalls, draft a proper skill, and keep you in the approval loop before anything is installed.

Detect

Find patterns across completed task history.

Distill

Ask an agent to draft a useful SKILL.md, not just a prompt snippet.

Approve

Review, edit, enable, disable, or delete from the Skills tab.

~/.agentforge/skills/release-check/SKILL.md

# Recurring release checklist distilled from runs
name: release-check
description: Verify package build, backend binary,
and Electron resources before a DMG release.

steps:
  - run frontend-quality gate
  - compile Bun backend
  - inspect Forge output

Bun and Electron, cleanly separated

A desktop app with a local service core.

Electron owns the native shell. The Bun backend owns scheduling, SQLite, agent execution, channels, heartbeats, and skills.

Electronmain + preload

Reactrenderer UI

Bun.serve127.0.0.1:9712

SQLite~/.agentforge

Remote control, local execution

Start work from the places you already talk.

Optional bridges can create and monitor tasks while AgentForge still runs the work locally on your Mac.

Send prompts to the bot, get task links and result summaries back.

Slack

Socket Mode events for task creation and status updates in channels.

Feishu / Lark

Long-connection support without exposing your machine to the public internet.

WeChat

Experimental local bridge for messaging-driven task creation.

Quick Start

Install it, point it at a repo, start delegating.

AgentForge is MIT licensed and runs as a local macOS app.

Download

Grab the latest AgentForge DMG from GitHub Releases for Apple Silicon Macs.

Install

Drag AgentForge into Applications. The app bundles its compiled Bun backend.

Run agents

Choose a working directory, select Claude Code or Codex CLI, and schedule the task.

Build from source

# Clone and run the Electron app
git clone https://github.com/hetaoBackend/agentforge.git
cd agentforge/taskboard-electron
bun install --ignore-scripts
bun run start

# Build the distributable DMG
bun run make