Heavy alpha · public preview

Orchestrate every agent CLI.

Steer multiple agents from one UI. Fan work across the providers each is best at. Use the subscriptions you already pay for. Push the compute off your laptop into cloud sandboxes.

Open the app→See how it works
Works with Claude·Gemini·Codex·Aider·Amp·opencode
claude
gemini
codex
opencode
🪢

Multi-agent orchestration

One card, six subtasks, three providers in parallel. Claude writes the schema, Gemini researches the index strategy, Codex stubs the test fixture. A coordinator LLM merges their outputs into one summary.

💳

Your subscriptions, not API spend

AgentOrc spawns the CLIs you already use. Runs consume your Claude Max / Gemini Advanced / Cursor quota — not pay-per-token API spend. Do more with what you're already paying for.

☁️

Give your laptop back

Optional rexec runtime executes agents in ephemeral cloud sandboxes. Your machine stays cool while a fleet grinds through your backlog. Docker preview for what they built.

Drives the CLIs you already use
Claude CodeCodexGemini CLIAiderSourcegraph AmpopencodeCursor+ customClaude CodeCodexGemini CLIAiderSourcegraph AmpopencodeCursor+ custom

You're already running six terminals.
There's a better way.

The state you're in

  • You pay for Claude Max and Gemini Advanced and ChatGPT Plus.
  • Three terminal tabs open, each running a different agent.
  • Copy/paste prompts between them. Forget which task is where.
  • Each one wedges your laptop with a 4GB Node process.
  • You wrote a shell script to coordinate them. It works okay.

The state AgentOrc gets you to

  • One UI. Threads, board, wiki, runs all in view.
  • Pick the agent per subtask. Hit fan-out. Watch three providers grind in parallel.
  • Run in a sandbox so your laptop fan isn't a jet engine.
  • Coordinator agent reads every output and writes the synthesis.
  • You steer. They build. You merge.

One card. Three providers. Parallel.

Live preview of what a multi-agent fan-out looks like inside AgentOrc.

card · migrate orders table without downtime
scroll to start demo…

How it works

Three steps from prompt to merged PR. No new CLI to learn — AgentOrc drives the ones you already use.

1

Brain-dump the task

Type what you want done into a board card. An LLM scaffolds it into a title, brief, subtasks, KPIs. Edit inline. Saves automatically.

2

Pin a provider per subtask

Each subtask gets a per-provider dropdown. Claude for the architecture, Gemini for the docs, Codex for the test stub. Or leave them on default for a single-agent run.

3

Fan out — watch them work

Hit the Fan Out button. AgentOrc groups subtasks by provider, spawns one CLI per group in a per-thread git worktree. A coordinator reads every output when they finish and posts one unified summary.

What's in the box

Not a thin wrapper. A real surface that takes agent orchestration as seriously as the agents themselves.

orders.go · branch +ship/8
42
43
44
45
46
47
func Migrate() {
+ if upgrade {
+ return nil
+ }
apply(ctx)
}

Inline code editor

CodeMirror surface for ~30 languages. Open files the agent edited, tweak inline, push the branch.

http://localhost:5173live

Live runtime preview

Docker container boots with your code. Dev server starts. Hot reload fires on file changes. Logs + preview iframe in one panel.

A∑

Multi-agent fan-out

One card, per-subtask provider. AgentOrc spawns parallel runs and a coordinator merges the outputs.

Graph view

Force-directed map of every thread, card, wiki page, run — and the edges between them. Hover to highlight neighbors.

Threads with inline runs
Per-thread persona + model. Agent runs render in the chat timeline.
Board with phases + deps
Kanban / sprint / phases / custom. Cards block + are blocked by each other.
Wiki that compounds
Agents auto-distill completed runs into wiki pages. Context grows per workspace.
Auto-fallback
Picked rexec but sandbox can't install the CLI? Falls back to exec transparently.

Why not just…

Just use Claude Code

Great single-agent CLI. But you're stuck with one provider, one model, and your laptop's RAM. No multi-agent fan-out. No history surface across sessions.

Use a shell script

Everyone writes one. They fall apart the moment you need history, retries, per-thread context, a UI, or coordination across providers. You're rebuilding AgentOrc badly.

Use Cursor / Windsurf

Editor-first IDEs with one inline agent. Different shape — they want you typing IN the editor. AgentOrc is the orchestration layer ABOVE: planning, fan-out, sandboxes.

Questions

Is it free?
Yes, source-available, runs on your hardware. You pay for whatever provider subscriptions / API keys you wire in. opencode is bundled as the recommended no-API-key entry path.
Does it need cloud / a SaaS account?
No. Single Go binary + Postgres on your machine. Cloud sandboxes (rexec) are opt-in; everything works locally without one.
Will it use my Claude Max / Gemini Advanced quota?
Yes — AgentOrc spawns the CLI you already have logged in. Runs consume your subscription quota, not pay-per-token API spend.
Which agents are supported?
Claude Code, Codex, Aider, Gemini CLI, Sourcegraph Amp, opencode — out of the box. Custom presets accept any shell command that reads from stdin.
Does it work with my repo?
Yes. Point a workspace at a local git repo path (or paste a URL and let AgentOrc clone it). Each run gets a per-thread git worktree. Push branches from the inline editor.
What's the moat vs the next person who builds this?
Depth across the surface: per-thread worktrees with CLI session resume, multi-provider coordinator synthesis, Docker runtime + preview, embedded editor with file-change hot reload, auto-fallback. None of those are 'one weekend' features.

Stop babysitting six terminals.

One UI, every provider, off your machine. Built for people who already pay for the agents — and want them to do more.

bash
scroll to start…
Open the app→View on GitHub