Skip to content
ReviewRamble

The weekly brief

AI, reviewed, in your inbox.

  • New reviews of tools, models, MCP servers and repos, with the verdict up front
  • Re-checked prices and ratings when something changes
  • New analysis pieces as they are published

Models, agents, MCP servers and the tools actually worth your time.

One email a week. Unsubscribe with the link in any brief. How we use your email

SwarmForge

Robert C. Martin's tmux-and-worktree harness that runs two to six role-based coding agents passing committed work down a pipeline.

Reviewed by ReviewRamble editors · Updated

Share this page:XLinkedInFacebookRedditWhatsAppEmail

Our rating

3.3/5

SwarmForge is a well-argued experiment in running coding agents like a disciplined team, but its missing licence and permission-bypassing defaults keep it a thing to study rather than adopt.

What is SwarmForge?

SwarmForge, from Robert C. Martin (Uncle Bob), runs several AI coding agents on one project, each in its own git worktree and tmux session. Agents do not talk to each other directly: they commit work and pass it on through durable handoff files, delivered by a small daemon, while a local web dashboard lets the operator start work, watch each agent, approve gates, answer questions and stop the swarm.

The main branch holds the shared runtime, written in Babashka (a Clojure scripting runtime), plus a shared "constitution" of engineering rules that every agent reads. Products live on other branches: a two-pack (coder, cleaner), a four-pack (specifier, coder, refactorer, architect), a six-pack that adds hardening and QA, and two multi-project "forge" setups with a planning lieutenant agent. Supported backends are Claude Code, Codex, GitHub Copilot and Grok.

Is SwarmForge worth it?

SwarmForge encodes one engineer's process as a pipeline of agents. In the six-pack, a specifier writes Gherkin scenarios and QA procedures, a coder implements, a cleaner refactors, an architect manages module boundaries, a hardener runs mutation tests and a QA agent follows the written procedures. The repository has about 4,100 GitHub stars and 411 forks since April 2026.

The handoff design is the interesting part. Each role works in its own worktree and passes committed work forward, so one agent cannot overwrite another's uncommitted files, and the shared rules come from the main branch so a pack cannot quietly replace them. The dashboard binds to 127.0.0.1, which keeps it off the network.

The defaults deserve a careful read. In the launcher, Claude Code starts with --permission-mode bypassPermissions and Codex and Copilot with --yolo unless your configuration already sets them, so every agent runs shell commands without asking. Each worktree is a full checkout with whatever credentials your environment holds. Run it in a container or a throwaway machine, not on a laptop with production keys.

The catch: there is no licence file, which under default copyright means you may read the code but have no written right to copy, modify or ship it. Activity has also slowed: the last commit on any branch was 7 September 2026. Installation is a curl of a shell helper onto your PATH, and it needs zsh, tmux and Babashka. The first line of the README warns that a SWARM crypto token is unrelated to the project.

Our take: as a reference for how to split agent work into roles with explicit handoffs, SwarmForge is worth an afternoon of reading; as a tool for a company codebase, Gas Town or Claude Squad come with a licence and fewer sharp edges.

Where it earns its place

  • A senior engineer evaluating agent workflows for a ten-person team

    See whether separate spec, code, refactor and QA agents produce cleaner output than one agent

    Run `get-swarm-forge six-pack` in a scratch copy of a side project inside a VM, then `./swarm`, and compare the resulting tests and structure with a single-agent run. Six agents means roughly six times the token spend of one.

  • A solo developer on a small Clojure or JavaScript project

    Get a second agent to clean up after the first one without manual prompting

    The two-pack runs a coder and a cleaner; the cleaner receives each committed change, refactors it and hands it back, while the dashboard shows both panes. Cost is two agent subscriptions or API budgets.

Before you commit

Setup
Install zsh, git, tmux and Babashka, curl the `get-swarm-forge` helper onto your PATH, run it with a pack name in your project, then `./swarm`.
Integrations
Claude Code, Codex, GitHub Copilot and Grok as agent backends, configured per role in swarmforge/swarmforge.conf.
Data and privacy
Runs locally; code and prompts go to whichever agent backends you configure. Agents run with permission prompts bypassed by default, and runtime state sits in .swarmforge/ in the project.
Cost at scale
The software costs nothing; each role is a separate agent session, so a six-pack consumes roughly six agents' worth of tokens or seats.
Limits
No licence; no tagged releases; macOS or Linux shell assumptions (zsh, tmux); last commit 7 September 2026.

What’s good

  • Two-, four- and six-agent packs with clear role boundaries
  • Committed handoffs between worktrees stop agents clobbering each other
  • Local dashboard for approvals and clarifications, bound to 127.0.0.1
  • Works with Claude Code, Codex, Copilot and Grok

What’s not

  • No licence file, so no right to reuse or modify the code
  • Agents start with permission prompts bypassed by default
  • No commits since 7 September 2026
  • Needs zsh, tmux and Babashka; installed by curl onto PATH

Best for

Engineers studying how to structure multi-agent coding pipelines, who can run it in an isolated machine and do not need to redistribute it.

SwarmForge alternatives

SwarmForge: common questions

Is SwarmForge worth it?
SwarmForge is a well-argued experiment in running coding agents like a disciplined team, but its missing licence and permission-bypassing defaults keep it a thing to study rather than adopt. We rate it 3.3 out of 5.
How much does SwarmForge cost?
Free to download, but the repository has no licence file, so reuse rights are not granted. You pay for the agent backends (Claude, Codex, Copilot or Grok) it runs.
What are the downsides of SwarmForge?
No licence file, so no right to reuse or modify the code; agents start with permission prompts bypassed by default; no commits since 7 September 2026; needs zsh, tmux and Babashka; installed by curl onto PATH.
Who is SwarmForge best for?
Engineers studying how to structure multi-agent coding pipelines, who can run it in an isolated machine and do not need to redistribute it.
How much does SwarmForge cost for a team or in production?
The software costs nothing; each role is a separate agent session, so a six-pack consumes roughly six agents' worth of tokens or seats.
Is SwarmForge safe for company data?
Runs locally; code and prompts go to whichever agent backends you configure. Agents run with permission prompts bypassed by default, and runtime state sits in .swarmforge/ in the project.
What are the best SwarmForge alternatives?
The closest alternatives we have reviewed are Gas Town, Claude Squad and Vibe Kanban.

Answers checked Nov 16, 2026.