Solo.io Open-Sources agentdesktop for Governing AI Agents on Developer Devices

Summary

Solo.io has released agentdesktop, an Apache 2.0 open-source project intended to govern AI-agent tooling running on developer laptops and desktops. It is designed to complement mobile-device management rather than replace it: existing MDM continues to manage the device and operating-system posture, while agentdesktop manages coding agents, MCP servers, skills, models, and other locally configured AI tools. Its inventory associates those tools with a device and user, and its control plane can version native configurations for supported agent harnesses. A central policy can, for example, permit writes to selected directories, block access to ~/.ssh, and restrict network access to approved domains; agentdesktop is intended to translate that policy into the JSON, JSONC, TOML, or other syntax required by each harness. The project also addresses long-lived provider API keys and OAuth tokens by allowing an endpoint daemon to authenticate a user and device through SSO, obtain a short-lived JWT, and have an LLM gateway inject the provider credential without storing it on the workstation. Gateway telemetry can attribute model and token usage to the user, device, and client label. Solo.io acknowledges that the client label is asserted rather than cryptographic proof of the executable, so another process in the same user boundary could potentially request a permitted label; future support for SPIFFE SVID identity is planned. Agentdesktop can run in standalone mode with local YAML and dry-run reconciliation, or in an enterprise deployment with an MDM-installed daemon and a centrally managed controller for enrollment, versioned configuration, credential issuance, and reporting. The controller can integrate with an existing identity provider, PKI, database, and LLM gateway, including on Kubernetes. MCP connections are inventoried but are not automatically proxied through agentgateway. At the time of the article, the planned GitHub code was not yet available, and the project’s usefulness will depend on keeping pace with rapidly changing tool schemas, credential systems, and sandbox behavior.