SwarmForge
github.com/unclebob/swarm-forge- Category
- AI Agents
- Rank
- No. 1602Tools index
Previous survey · No. 1597 ·
- Pricing
- Open Source
- Type
- TOOL
- Use case
- Agent Building · Coding
- Interfaces
- CLI
- Builder
- unclebob
- GitHub
- 3.9k stars
- Date
About
SwarmForge is a tmux-based orchestration platform that coordinates multiple AI coding agents (Claude, Codex, Copilot, Grok) working in separate git worktrees on the same project. It assigns role-based prompts (coder, cleaner, architect, QA, etc.) and a shared 'constitution' of engineering rules across configurable workflows (two-pack, four-pack, six-pack) to manage handoffs between agents without conflicts.
What it does
SwarmForge runs a local development process as a set of visible terminal workers. Work moves through durable inbox files, with priorities, task state, and completion history recorded on disk. Each worker follows project rules and passes committed changes through a chosen sequence of implementation and quality checks.
Why it's ranked here
The strongest idea is not simply running several agents. It gives their coordination explicit state, validation, ordering, and restart history. The repository also tests configuration parsing, task and batch queues, handoff validation, daemon shutdown, and terminal behavior. That discipline makes the approach credible, though its many local dependencies and branch-based distribution add operational friction.
What's good
Handoffs use validated headers, fixed commit abbreviations, known recipients, atomic queue writes, and priority ordering. Task and batch modes support different review roles without changing the transport. Completed, failed, sent, queued, and active work remain visible on disk. Existing project rules can override shared defaults without startup overwriting them.
Tradeoffs
Setup requires zsh, Git, tmux, Babashka, and at least one separately configured coding agent. Runnable configurations live on dedicated branches rather than the main branch. The workflow creates several worktrees, sessions, terminal windows, and local state directories. Manual sleep or display locking can still interrupt workers, and the full process may be excessive for small edits.
How to use it well
Use it for repository work where several coding agents need explicit ownership, review stages, and auditable handoffs. Start with the compact implementation and cleanup flow, then choose specification or full QA stages when project risk justifies them. It does not supply an AI model, remote execution service, or general project tracker. You must bring a configured agent backend.
Technical notes+
README.md defines the documentary main branch, runnable workflow branches, prerequisites, bootstrap archive flow, and local state model. swarmforge/scripts/swarmforge.bb parses role configuration, validates supported backends and prompts, prepares tmux state, initializes Git when needed, and creates worktrees. swarmforge/scripts/handoffd.bb polls role outboxes, adds delivery headers, writes recipient inbox files, sends generic tmux notifications, and archives sent or failed work. swarmforge/scripts/swarm_handoff.bb validates drafts, recipients, priorities, message types, tasks, and ten-character commit abbreviations. swarmforge/scripts/handoff_lib.bb manages role lookup, headers, queue display, and locked sequence numbers. bb.edn runs the Clojure tests in test/swarmforge/handoff_test.clj and test/swarmforge/script_test.clj. close-swarm and swarmforge/scripts/swarm-cleanup.sh stop sessions, the daemon, and tracked terminal windows.
Observed
- Primary implementation
- Babashka Clojure scripts with zsh wrappers
- Interface
- Local command-line launcher and handoff helper commands
- Installation surface
- Runnable branch archive copied into the target project
- Required tools
- zsh, Git, tmux, Babashka, and a configured agent backend
- Supported agent backends
- Claude, Codex, Copilot, and Grok
- Runtime state
- Project-local files store sessions, queues, delivery history, and daemon state
- Testing
- Clojure tests cover launcher parsing, handoff queues, validation, cleanup, and terminal behavior
Read from README.md, bb.edn, close-swarm, swarmforge/handoff-protocol.md, swarmforge/scripts/handoffd.bb, test/swarmforge/script_test.clj, swarmforge/scripts/swarmforge.bb, swarmforge/scripts/swarmforge.sh, test/swarmforge/handoff_test.clj, swarmforge/scripts/handoff_lib.bb, swarmforge/scripts/swarm-cleanup.sh, swarmforge/scripts/swarm_handoff.bb.
What it can do
Assign role-based prompts to agents (coder, cleaner, architect, QA)
Role definitions → Configured agent roles
Apply a shared 'constitution' of engineering rules across agents
Constitution prompt files → Governed agent behavior
Select configurable multi-agent workflows (two-pack, four-pack, six-pack)
Workflow selection → Launched agent swarm
Tags
Comments (0)
No comments yet
Editorially curated, with community endorsements as a secondary signal. Corrections welcome.