Swarmaa gives Claude Code, Codex, Antigravity, Dot and Grok one shared memory and one task queue, and sends each task to whichever agent has the most quota left. It runs on your Mac.
Works with the agents you already pay for
No new agent to learn. Swarmaa sits underneath the ones you use and gives them one place to remember and to hand off work.
Run npm run setup. It gives each agent its own scoped key and registers Swarmaa over MCP in Claude Code, Codex and Antigravity.
Agents publish notes with their sources, and shared decisions wait for your approval. Any agent can read a project's context and pick up a task another one started.
Workers report each provider's usage and reset time. Submit a task as auto.read and it goes to the live agent with the most headroom.
Memory, durable tasks, routing and guardrails, in one small service that starts with your Mac.
Working with several agents means a lot of invisible work done by hand. Watch the waterline drop as Swarmaa takes it over, until only the decisions are left to you.
Start where switching between agents costs you the most.
Submit a review and let Claude and Codex workers read the project in parallel. Results land as artifacts any agent can read.
Command-line agents connect over MCP, each with its own scoped key. Hosted agents reach the same memory over HTTPS through a tunnel you control. Agents share only what they publish.
Swarmaa is in private beta. Join the waitlist and we'll tell you when the portal opens.
No. It's the memory and dispatch layer underneath the agents you already use. Swarmaa never calls a model itself; your agents do the work on your existing subscriptions.
On your Mac, in a local PostgreSQL database. Each agent has its own scoped key in an owner-only file. Hosted agents like Dot and Grok reach it only through a tunnel you set up.
No. Shared memory holds what agents publish through Swarmaa, with sources attached. It doesn't import chat histories or any vendor's built-in memory.
Background workers are read-only by default. Claude's file-edit mode is a separate opt-in per project, and no task can choose which program, folder or command runs.
Workers read each provider's usage windows from the CLI's own login every few minutes and report them. A task submitted as auto.read goes to the live worker with the most headroom. If no fresh reading exists, any live worker can take it, so work never stalls.
A place on swarmaa.com to see and manage your swarms from any device. It's being built now; join the waitlist to hear when it opens.