swarmagent.dev
GitHub
DocsModelsResourcesPricingNewsContactGitHub

Use casesParallel coding agents

Run coding agents in parallel on your own server

Install Swarm on a Linux machine and connect it to Claude Code, Codex or ChatGPT. Your AI hands each job to Swarm and gets a task id back at once, so several jobs run side by side. Each one works on its own branch and comes back with a report, a diff and what it cost. You keep working and merge what you like.

Why one chat is not enough

A coding assistant in one terminal does one thing at a time, and it stops when you close the laptop. Real work is a queue: a failing test, a dependency bump, a small feature, a refactor. Swarm takes that queue off your terminal. Your AI stays the one you talk to; Swarm is where the jobs run.

How it works

  1. Install Swarm on an Ubuntu 24.04 x86-64 machine: your PC, a VPS or a cloud VM. Your AI does it: curl -fsSL https://swarmagent.dev/install | sudo bash.
  2. Connect your AI. Claude Code gets the Swarm plugin; Codex and other clients get Swarm's MCP server at 127.0.0.1:7783/mcp.
  3. Pick how models are paid for: your ChatGPT plan, your Claude plan through Claude Code, or an API key (OpenAI, Anthropic, Google, OpenRouter, Fireworks).
  4. Give it jobs. Each swarm_run_task with the coder role returns a task id straight away. Start as many as your limits allow; your AI waits on them with swarm_wait_task.
  5. Review and merge. swarm_task_result returns the report and the diff; swarm_integrate_task merges it. Not right? swarm_send_message sends a correction to the same task, on the same branch.
Say this to your AI

Fix the flaky upload test, bump the Stripe SDK, and add CSV export to the reports page. Run them in parallel and show me each diff.

  • swarm_run_task × 3
  • swarm_wait_task
  • swarm_task_result

Three coder tasks are running, each on its own branch. The test fix is ready: +18 −6, $0.09. The other two are still working.

What keeps parallel jobs from colliding

  • Every job works on its own branch. Nothing reaches your folder until your AI integrates it.
  • Limits you set (swarm_set_limits) cap how many jobs run at once and what one job may spend.
  • Each result reports its tokens and estimated cost, so the queue never surprises you.
  • Add a reviewer step with a workflow (swarm_set_workflow) and every finished job is checked against its brief before it comes back, or sent back with notes.

Where the agents run

On your machine, as Swarm's own user. By default an agent's commands run without permission prompts and without a sandbox; install.sh sandbox on runs every command in one container per project, and an environment gives a folder its own toolchain container with its secrets as stand-ins. Model requests still go to the provider you connected.

Questions

How many coding agents can run at once?

As many as the limits you set for the machine allow: Swarm caps concurrent jobs and per-job spend with swarm_set_limits. The practical limit is your machine and your model plan.

Does this work with Claude Code?

Yes. Claude Code connects to Swarm through the Swarm plugin (or as an MCP server) and hands jobs to it. Swarm's agents can also run on your Claude plan through Claude Code on the Swarm machine.

Does it work with Codex or ChatGPT?

Yes. Codex connects to Swarm's MCP server, and Swarm can run its agents on your ChatGPT plan after a one-time sign-in. ChatGPT itself reaches Swarm through OpenAI's tunnel.

What does it cost?

Swarm is free and open source (Apache 2.0). The models run on your own plan or API key; every job reports its token use and estimated cost.

Can it run on a Mac?

Not yet. Run Swarm on an Ubuntu 24.04 x86-64 server and drive it from your Mac; your AI can install it over SSH.

Updated 24 October 2026