Software details
Description
Abbreviation
Latest version
Authors
Organization
Programming language
Operating system
Genre
License
Website
Repository
Status

OpenMule is an open-source proposal for a decentralized marketplace platform designed to facilitate transactions between end-users and service providers of automated agents. It positions itself as a free and open-source software alternative to Mulerun, enabling the discovery, purchase, and management of tasks performed by both digital AI agents and physical automation systems.[3] The project emphasizes a crypto-first payment system, security, and a vision for a future distributed network called "MuleNet." The first public announcement and repository activity appeared in September 2025, with a self-described "vibe-coded-marketplace" release tagged on February 25, 2026.[1]

Overview

OpenMule aims to create a global marketplace where humans and automated agents can collaborate and transact. Its core vision is to become a "Freelancer.com for AI," supporting a wide range of services from simple digital tasks to complex physical hardware control. The project description on its main website calls it "a distributed global marketplace for relaying agent traffic to computer endpoints" that is meant to be "distributed, trustless, and user-friendly."[3]

The author, James Brown (GitHub username james4ever0, affiliated with Cybergod AGI Research), circulated the design as a collaboration proposal across adjacent open-source projects in September 2025.[5][6] Issues titled "Proposal for Collaboration: Project OpenMule" were opened against the Chisel TCP tunneling project and ZeroTier, the peer-to-peer virtual network platform.[5][6] The ZeroTier maintainers labeled the issue "Nope" and declined, but the documents remain useful primary sources for the design intent.[6]

Motivation

The project was proposed to address the limitations of closed-source, centralized agent marketplaces. The author argues that as more economic activity moves through agent-driven workflows, the underlying routing, billing, and trust layers should not be controlled by a single vendor.[5] Key motivations include:

Providing an open-source platform that allows for community contribution and modification.
Enabling users to retain control over their data and execution environments by allowing self-hosted agents.
Creating a transparent, crypto-native economy for agent services.
Supporting not just digital AI agents but also physical automation and robotics, including the ability to drive consumer hardware like game consoles and personal computers.
Integrating Human-in-the-Loop (HITL) workflows, allowing humans to supervise, correct, and earn money alongside AI.

The positioning as an alternative to Mulerun is explicit. Mulerun is a commercial AI agent platform marketing dedicated cloud computers that run agent workflows 24/7.[7] OpenMule borrows the "Mule" naming and the underlying premise of an always-on agent workforce, but recasts the platform as a federated, crypto-native marketplace rather than a single hosted service.[3] The OpenMule organization on GitHub also maintains mulerouter-skills, a Claude Code plugin that exposes the MuleRouter and Mulerun multimodal generation APIs through agent skills.[4]

Technical architecture

OpenMule is designed with a modern, scalable tech stack to support real-time interaction and complex agent workflows. The GitHub repository reports a polyglot codebase weighted toward HTML, Python, Go, and Rust, reflecting the project's split between a static documentation site, the FastAPI-based control plane, and the lower-level relay and tooling components.[1]

Component	Technology
Frontend	Vue 3
Backend	FastAPI, WebSocket (for real-time control)
Agent framework	TryCUA/CUA (compatible)
AI billing and metering	LiteLLM
Configuration	YAML (for agent prompts, Dockerfiles, billing, credentials)
Deployment	Docker, self-hosted servers, cloud instances
License	The Unlicense

Key components

Marketplace relay: Acts as a secure intermediary, routing traffic between clients and agents running on provider-owned or OpenMule-hosted servers. The relay is the connective tissue that lets a buyer in one location drive a worker machine in another without exposing either side's full intranet.[5]
Escrow and payments service: Manages crypto-first transactions, holding payments in escrow and releasing them upon task completion or milestones. The design supports flat-rate, subscription, and bid-based pricing, plus a minimum wage-per-minute rate that compensates a worker for time spent on a task even before the final deliverable lands.
Safety and control web UI: Provides a web interface for users to monitor controlled machines, view logs, configure settings, and trigger emergency stops. The UI is password protected to limit access to the operator of the controlled machine.
Connection health probes: Background checks that detect when a relay link or controlled endpoint becomes unhealthy, so a stalled or hijacked agent does not silently burn through escrowed funds.
Accounting framework: Internal billing and ledger code that draws on cybergod-gym, a related Cybergod AGI Research project, to track per-agent costs and reconcile them with payments held in escrow.

Agent framework choices

The project is built to be compatible with TryCUA's CUA framework, an open-source toolkit for Computer Using Agents that drives a virtual machine through screenshots and synthetic input.[5] By aligning with CUA, OpenMule inherits a body of work on sandboxed agent execution and gains a clear answer to the question of how a marketplace lists a unit of work: a CUA agent ships as a container plus a YAML descriptor for prompts, billing, and credentials, and the marketplace can compose, price, and route it like any other listing.

LiteLLM is used as the AI billing and metering layer. LiteLLM is an open-source proxy that exposes a single OpenAI-compatible API across more than 100 model providers and tracks token usage, costs, and rate limits per request. In OpenMule this gives providers a uniform way to meter the language-model calls their agents make and pass the cost through to the buyer in a verifiable way.[5]

A potential integration with the Model Context Protocol (MCP) is also called out in the proposal documents.[5]

Features

Marketplace functionality

Service listings: Providers can list agents with descriptions, capabilities, and pricing (fixed-price or subscription).
Task bidding: Users can post tasks and receive bids from agents for negotiable, versatile work. A buyer who is not sure exactly what they need can describe an outcome and let providers compete on price and approach.
Reputation system: Clients can rate and comment on service providers to build trust. "Certified" verification badges for both providers and users are planned, with the goal of separating long-running reliable operators from one-off accounts.
Dispute resolution: A system for handling refund requests, potentially involving community validators and automated agent review. Relay nodes can also act as third-party arbiters and earn fees when their resolution is accepted by both sides.
Minimum wage per minute: During execution, the controlled machine is paid a baseline rate for time actively occupied, with milestone bonuses released as work progresses. This is meant to keep providers honest about long-running jobs and to make abandoned tasks visible.[5]

Agent deployment and control

Agents can be deployed on user-defined remote machines, cloud instances, OpenMule-hosted servers, or directly on end-user devices.
Supports control of physical hardware, including game consoles (Xbox, PS5), personal computers, and robots via high-level APIs. Driving consumer hardware is unusual for an agent marketplace and is one of the more distinctive parts of the proposal.[5]
Allows composition of multiple agents to handle complex workloads. A buyer can chain a research agent, a planning agent, and an execution agent into one workflow.
Supports read-only observation modes and randomized testing agents, useful for benchmarking and for letting auditors inspect a workflow without granting full control.
Lets human workers compete on the same marketplace as AI agents, with the same payment and dispute mechanics applied to both. This is the explicit Human-in-the-Loop economy the project advertises.

Security protocol

Security is a paramount concern, especially for bare-metal machine control, where a misbehaving agent could trash a real machine instead of a disposable virtual one. The proposal documents lay out a layered approach:

Network isolation: Eliminates direct intranet access to user machines. The relay is the only public ingress point, so the controlled machine does not need to expose RDP, VNC, or SSH ports to the internet.
Virtualization recommendation: Users are strongly advised to use sandboxed virtual machines for any agent that controls a desktop session. Running an unknown agent on a bare metal personal computer is treated as a power-user mode with extra warnings.
Bare-metal emergency stop: On non-virtualized machines, unexpected user input such as keyboard or mouse movement triggers an immediate emergency stop of the agent, on the assumption that the human owner has walked back to the machine and wants control returned.
Operation logging and developer identification: Every controlled session is recorded locally and tied to a developer identity, so disputes can be reviewed against a concrete log rather than self-reported claims.
Local recording for privacy: Session recordings are kept on the controlled machine by default, not centralized, to limit how much data the marketplace operator can see about any individual buyer or worker.[5]

Payment system

The payment design is the most opinionated part of the proposal. OpenMule is described as crypto-first rather than crypto-only: cryptocurrency rails are the default because they are programmable, but the system is expected to accept fiat through gateways once the network has volume to justify the integration work.[5]

A single task can mix several payment modes:

Fixed price for well-defined, repeatable work.
Subscription for ongoing access to an agent, similar to a managed SaaS plan.
Bid-based for negotiable jobs where buyers and providers settle on a price up front.
Per-minute floor while the worker is actively engaged, so providers do not lose money on long jobs that fail late.
Milestone releases that drip escrowed funds out as the work clears intermediate checkpoints.

Every network participant earns: relay operators, escrow services, environment providers (the people whose machines the work runs on), and validators who resolve disputes.[5] The intent is that the operator of a relay or arbitration node is paid in the same crypto economy as the agents themselves, which is what is supposed to make the network self-sustaining once it grows past the initial free tier.

Development status

As of early 2026, OpenMule remains an active open proposal seeking contributors. The GitHub repository at James4Ever0/openmule reports roughly 89 commits, a small number of stars and forks, and two tagged releases, the most recent of which is labeled vibe-coded-marketplace and dated February 25, 2026.[1] The mirrored repository on GitLab at openmule/openmule was created on September 13, 2025 and has roughly 69 commits on the main branch.[2] Both repositories use The Unlicense, a public domain dedication, which is unusual for a project that hopes to anchor a multi-party economy and which any future enterprise adopter would likely want to revisit.

The official website at openmule.netlify.app is a small Netlify-hosted static site. As of the most recent snapshot it carries a single dated blog post (September 15, 2025) titled "OpenMule: An Open Source MuleRun Alternative," which functions as the announcement of the project.[3]

The project is actively recruiting contributors across several domains:

Distributed systems and network relay engineering
AI agent development on the TryCUA/CUA framework
Physical KVM and hardware control, including driving consumer game consoles
Blockchain development for escrow, payments, and validator incentives
Frontend (Vue.js) and backend (FastAPI) development
UI/UX design for the safety and control web interface

Contact channels listed on the site include a foxmail address and a Discord invite, alongside the GitHub and GitLab issue trackers.[3]

Future vision: the MuleNet

The long-term goal is to evolve from a single hosted marketplace into a fully decentralized network the proposal calls MuleNet.[5] The design sketch covers three layers:

Distributed relay framework: Building MuleNet on top of a blockchain for decentralized service discovery, routing, and payment escrow. Instead of one OpenMule company running every relay, anyone can stand up a node and route traffic for fees.
Open protocol: Creating an open standard so that any vendor can build compatible clients and servers and still interoperate with the rest of the network. The goal is to make the marketplace economy itself the product, rather than any one implementation.
Node incentives: Operators of relay or validator nodes earn fees for providing services to the network, including dispute resolution and traffic routing. Validators that consistently produce accepted rulings earn more, which is meant to filter for reliable arbiters over time.

Whether MuleNet ever ships in the form described is genuinely uncertain. The project is one person plus collaborators, the codebase is small, and several of the proposed integrations (Chisel, ZeroTier) were declined by the upstream maintainers.[6] The proposal documents are useful as a snapshot of what an open MuleRun-style marketplace might look like, but the gap between proposal and production-grade decentralized network is large.

Reception and context

OpenMule arrives during a broader push toward agent marketplaces. Academic groups such as the Berkeley Gorilla project have argued that an open agent marketplace is a natural successor to today's app stores.[8] The most direct commercial comparison is still Mulerun itself, which sells dedicated cloud computers running an agent workforce on subscription pricing rather than a federated open protocol.[7] OpenMule's bet is that the relay, escrow, and reputation layers will end up looking more like email or BitTorrent than like a single SaaS product, and that an open-protocol marketplace is therefore worth building even if the first version is rough.

The collaboration proposals on the Chisel and ZeroTier trackers show the author treating existing tunneling and overlay-network tooling as candidate primitives for the OpenMule relay, rather than reinventing the wire protocol from scratch. The upstream maintainers did not pick up the collaboration, but the documents remain in place as part of the public design record for the project.[5][6]

See also

References

OpenMule project repository on GitHub. "James4Ever0/openmule: An open-source MuleRun." Retrieved May 2026. https://github.com/James4Ever0/openmule
OpenMule project repository on GitLab. "openmule/openmule." Created September 13, 2025. Retrieved May 2026. https://gitlab.com/openmule/openmule
Official OpenMule website. "OpenMule, MuleRun FOSS Alternative." Blog post dated September 15, 2025. Retrieved May 2026. https://openmule.netlify.app/
OpenMule organization on GitHub. "mulerouter-skills, Agent Skills For MuleRouter and MuleRun API." Retrieved May 2026. https://github.com/openmule/mulerouter-skills
James Brown. "Proposal for Collaboration: Project OpenMule, A Decentralized Marketplace for AI and Physical Agent Services." GitHub issue jpillora/chisel#569. Retrieved May 2026. https://github.com/jpillora/chisel/issues/569
James Brown. "Proposal for Collaboration: Project OpenMule, A Decentralized Marketplace for AI and Physical Agent Services." GitHub issue zerotier/ZeroTierOne#2508, opened September 11, 2025. Retrieved May 2026. https://github.com/zerotier/ZeroTierOne/issues/2508
Mulerun official website. "MuleRun, The AI Agent That Gets Work Done." Retrieved May 2026. https://mulerun.com/
Berkeley Gorilla project. "Agent Marketplace," research blog post. Retrieved May 2026. https://gorilla.cs.berkeley.edu/blogs/11_agent_marketplace.html

Editing: OpenMule