Launching today
AgentFleet OSS
One queue for Claude Code and Codex — on every machine
1 follower
One queue for Claude Code and Codex — on every machine
1 follower
Self-hosted scheduler for the AI coding CLIs you already have — Claude Code, Codex, Aider, Gemini, Qwen. One queue across your machines: a task goes to whichever device is free, and if a device dies mid-task it is re-queued and re-dispatched to another machine, with the partial output carried over on the local CLI path. Not a new agent framework, and not session-state migration. LAN-first with relay fallback. MIT.
Free
Launch Team
Render Sandboxes: a computer for every agentGive every agent its own machine. Starts in under a second
Promoted
Maker here 👋
I built AgentFleet because every agent CLI I already use lived on a different machine and none of them knew the others existed. Codex on the laptop, Claude Code on the home server, Cursor on the work box, Aider on a spare NUC.
Most multi-agent frameworks want you to rebuild agents inside their runtime. Remote-access tools just hand one task from device to device. AgentFleet treats each device as a first-class ship and coordinates them into one fleet — LAN-first, cloud relay as fallback.
What's actually measured (reproducible, scripts in the repo):
• 303 tasks/s scheduler throughput; 377 tasks/s at 1000 tasks
• 8 devices × 24 agents, 1000 tasks, a device killed mid-run → 1000/1000 completed, 0 lost. Running tasks migrate back to queued and get re-dispatched.
• Disconnect detection ~1s (was ~35s)
• One scheduling fix: 13 → 414 tasks/s (30×) after removing three O(n²) DB lookups from the dispatch loop
• Real cloud servers: 1 Ubuntu + 2 Windows, real codex/claude CLI, 5/5 completed with zero account login
Honest status: only the Grok adapter is fully stable today; the other ~12 are scaffolds — install/discovery/gateway routing work, direct headless CLI execution still needs per-platform testing. That's exactly where I'd love help.
What feedback would help most: if you run any of these CLIs across more than one machine, tell me what breaks first. Cross-device state is where the real bugs hide, not throughput.
Repo: https://github.com/narkco-kang/agent-fleet (MIT, Node v24 + TypeScript)