Myco Connects AI Agents Across Machines

What changed

Myco introduces a unified network for AI agents, aiming to solve the problem of agents operating in isolation. Previously, developers often acted as the manual bridge between agents running on different machines or in different environments, copying prompts and re-explaining context. Myco centralizes this by allowing agents to discover and interact with each other through a single hub. Key functionalities include:

Cloud Agent Access to Local Resources

Cloud-based agents, such as those on claude.ai or VPS instances, can now access resources on a developer's local machine. This includes local configurations, logged-in CLIs, build toolchains, or local databases. Myco enables these cloud agents to execute commands on a local shell node or hand tasks directly to a local Claude Code or Codex session. The execution happens within the local machine's existing permissions, with all commands recorded. Importantly, the local computer opens no ports, and access is restricted to agents explicitly allowed by the user. This feature supports Mac, Linux, and Windows environments, with commands executed via sh or PowerShell. Limitations include single command execution at a time, a default 60-second timeout, and a 64 KB output cap, with no direct file transfer or screen control.

Inter-Agent Task Delegation

Myco facilitates task delegation between agents, even those belonging to different users or running on separate machines. A developer's agent can send a task, along with necessary context, directly to a teammate's agent. The response is then returned to the originating agent. Access control is managed by agent owners, who define which other agents can send them tasks. For privacy, teams can enable end-to-end encryption for messages exchanged within the hub. All inter-agent exchanges are logged within the Myco application.

Unified Agent Network

Once agents join the Myco network, they can discover and interact with each other. This includes Claude Code and Codex sessions on local machines and servers, as well as cloud-based agents. The myco agent list command provides visibility into available agents, their locations, and online status. Agents can be organized with groups and tags. The system supports delegation, where one agent can assign tasks to another and await results.

Setup and Availability

Myco is distributed as a single Go binary, simplifying setup. It offers a graphical application for Mac (signed and notarized) and a preview version for Windows. For Linux servers, a command-line installation script is available. The setup involves starting a hub (either locally or on a server), adding nodes (agents) by providing a single-line command to the agent's session, and then initiating tasks through the application's interface or by instructing one agent to involve another.

Why it matters for builders

Myco significantly lowers the barrier to creating complex, multi-agent systems. Builders can now orchestrate workflows that span local development environments and cloud-based AI services without manual intervention. This allows for more powerful agent capabilities, such as leveraging local tools or accessing proprietary data securely, and enables collaborative AI development where agents from different team members can seamlessly work together.

Practical impact

Developers can now experiment with distributed AI architectures more easily. For instance, a cloud-based research agent could be instructed to run code analysis on a local machine's repository, or a team member's agent could query another's specialized data processing service. The simplified setup, with a single binary and one-line installation for hubs, encourages adoption. Builders should explore connecting their existing Claude Code and Codex instances to a Myco hub to test inter-agent communication and local resource access.

Caveats and source limits

The source material indicates that the Windows version is currently in preview and not yet code-signed. Shell nodes have limitations, including executing one command at a time, a default timeout of 60 seconds, and an output cap of 64 KB. There is no direct file transfer capability, nor screen or mouse control for remote machines. Codex cloud agents are mentioned as not running myco themselves, which may limit their ability to initiate calls to local machines. The provided source does not include independent benchmark results or detailed performance metrics for the agent networking capabilities.