agentdesktop

Sources: GitHub repository · Introducing agentdesktop (Solo.io blog) · Techstrong.ai coverage

Identity and role

Solo.io released agentdesktop on September 3, 2026 under an Apache 2.0 license, at github.com/agentdesktop-dev/agentdesktop. It is an endpoint governance layer for the AI agent harnesses developers run directly on their machines — Claude Code, Codex, and VS Code agent extensions — and for those harnesses’ MCP client configurations. It functions as the endpoint control plane complementing agentgateway, Solo.io’s server-side data plane: agentdesktop governs the harness on the desktop, agentgateway enforces policy on traffic in the cloud.

Governance mechanisms

agentdesktop discovers every agent harness and every MCP server registered in a harness’s configuration, building an endpoint-level inventory that surfaces shadow agent usage a fleet-wide security team otherwise cannot see. It translates centrally declared policy into each harness’s own native configuration format — sandbox filesystem and network restrictions expressed the way Claude Code or Codex natively understands them, not as OS-level rules — which distinguishes it from conventional mobile device management. It treats MCP servers as first-class policy objects, distinguishing authorized sources from unauthorized ones and enforcing that distinction at the point a harness would otherwise connect to one.

Identity, credentials, and observability

Each agent instance on a desktop is bound to device and user identity context, giving it a non-human identity distinct from the developer running it, and that binding drives least-privilege, just-in-time credential issuance: agentdesktop replaces distributed, long-lived API keys with short-lived credentials scoped to the specific agent, user, and device, injected at runtime through agentgateway rather than stored on the workstation. This is a desktop-scoped application of the credential proxy pattern and the practical mechanism behind NHI governance for AI agents at the endpoint — the harness itself never holds a standing secret. Session and tool-use telemetry is available on an opt-in basis, so an organization can monitor agent behavior and resource consumption without mandatory always-on collection.

Deployment modes

agentdesktop runs in two modes. Standalone mode provides local governance on a single machine. Fleet mode distributes the endpoint daemon through existing mobile device management, and a central controller integrates with the organization’s identity provider, public key infrastructure, and agentgateway for coordinated, organization-wide enforcement. The product functions as an endpoint-resident instance of the guardian agent pattern — software that supervises what other agents on the same machine are permitted to do — and its policy-enforcement and credential-issuance design bears on the CMM’s identity and egress domains.

Notable statements

Idit Levine, founder and CEO, Solo.io, on the sprawl the product addresses: “Every organization is struggling with AI agents and tooling sprawl. We at Solo.io are no different.” On where agents now run: “Laptops and desktops are the first real production environment for AI agents.”