Skip to content

Fleet architecture

How the fleet fits together.

How pooled machines, local models, relays and sessions are designed to fit together. OmniConflux is a desktop coding app for Omarchy, macOS on Apple silicon and Windows, version 0.3.0, in closed beta.

Illustration: three machines wired to a small glass ring, the relay, with a crystal core turning inside it; glass session capsules ride the links and four worktree branches fan out of the relay, one light each.
Fleet architecture diagramIn the design, four of your machines, a laptop, a desktop, a mini PC and a server, connect through a secure relay to a coordinator on your primary machine. The coordinator holds the task queue and rate-limit budget, sessions that can pause, resume and fork, one git worktree per agent, and a verifier from a different vendor.Laptopagents and vendor toolsDesktoplocal model on GPUMini PCbackground agentsServerlarge local modelSecure relayencrypted, machines dial outCoordinator on your primary machineholds the four parts belowTask queue andrate-limit budgetSessions: pause,resume, forkWorktrees: onebranch per agentVerifier from adifferent vendorFleet architecture diagramIn the design, four of your machines, a laptop, a desktop, a mini PC and a server, connect through a secure relay to a coordinator on your primary machine. The coordinator holds the task queue and rate-limit budget, sessions that can pause, resume and fork, one git worktree per agent, and a verifier from a different vendor.Laptopagents andvendor toolsDesktoplocal model on GPUMini PCbackground agentsServerlarge local modelSecure relayencrypted, machines dial outCoordinator on your primary machineholds the four parts belowTask queue and rate-limit budgetSessions: pause, resume, forkWorktrees: one branch per agentVerifier from a different vendor
In the design, your machines join through a secure relay, and the coordinator queues tasks, budgets rate limits, keeps sessions, isolates each agent in a worktree and sends finished work to a verifier from a different vendor.

Pooled machines

In the design, you install OmniConflux on each machine you want in the fleet and pair it once. Each machine reports what it can offer: CPU, memory, GPU memory, the agent tools installed on it and the local models it holds. One machine acts as the coordinator and hands out work to the machine that fits.

Local models

In the design, local models run through Ollama or llama.cpp on your own hardware. GPU memory decides which model fits where, so a large model lands on the server and a small one can run on a laptop. Local work is designed to stay on the machine it runs on.

Secure relay

In the design, machines reach each other over SSH or an encrypted relay. Each machine dials out, so a box behind a home router does not need an open inbound port. The relay carries fleet traffic between your machines. It is not where your work is stored.

Sessions: pause, resume, fork

In the design, a session is an agent process, its worktree and its transcript. Pause stops the process and keeps the worktree. Resume starts it again from the transcript. Fork copies the worktree to a new branch and starts a second agent from the same point, so you can try two approaches side by side.

Git worktree isolation

In the design, every agent works in its own git worktree on its own branch, no two agents write to the same checkout, nothing reaches your main branch before it has been reviewed, and merges happen one at a time.

Rate-limit coordination

In the design, agents that share one API key or one vendor login share one budget, as the vendor allows. The coordinator is designed to track that budget across the fleet and to spread requests out, so that a rate-limit error makes the fleet back off together instead of every agent retrying at once. If you allow it, queued work is designed to move to another vendor or a local model. This does not remove a provider's limits or terms.

  • Models, three ways

    In the design: the vendor's own tool on your machine under your login; your own API key, direct to Anthropic, OpenAI, Google or any other provider; or local models.

  • Credentials stay put

    In the design, API keys and vendor logins stay on the machine that uses them, and seat logins never pass through a hosted backend.

  • Open protocols

    Designed around MCP for tools, A2A and ACP for agent coordination, and OAuth 2.1 and OpenID Connect with PKCE for sign-in.

Keep reading

See how the verification engine is designed to review each change, and see an illustration of routing tasks by difficulty.

Questions

Which agents can run in the fleet?

In the design, any command-line agent can run at the process level, including Claude Code, Codex CLI, Gemini CLI, Aider and OpenCode. Agents that speak MCP or ACP are designed to get deeper control: state, handoff and approvals. Some agents need a short setup step.

How do subscriptions work across machines?

In the design, a subscription runs through the vendor's own tool, on your machine, under your login, as the vendor allows. OmniConflux does not pool personal accounts or share one seat across many cloud agents.

Do agents share context across vendors?

In the design, they share work through git and files. We do not claim lossless context transfer between vendors; each agent reads the repository state it is given.

What about a managed fleet?

Self-serve is the default. A managed fleet is planned, not live. If you want us to run machines for you, talk to us. It will run on your API keys or enterprise tokens, never on personal subscription seats.

All product names are trademarks of their owners. Naming a product describes designed compatibility, not endorsement.