OpenWork

Free desktop app with no account required The MIT-licensed desktop app runs on macOS, Windows, and Linux, works locally on authorized folders, and needs no OpenWork account for local use.[1][2][3]

More than 50 models plus local inference The AI Gateway fronts OpenAI, Anthropic, Google, Mistral, and other providers with one set of keys and limits, and local models are documented as an option.[1][2]

One-click reusable setups Skills, plugins, and MCP connections are packaged once and shared org-wide, so a teammate imports a working configuration instead of rebuilding it.[1][2]

MCP Gateway for the agents teams already use One documented command connects Claude Code, Cursor, or Codex to the org's skills and tools over HTTP.[1]

Browser automation and scheduled tasks The vendor's comparison table lists browser automation as included and scheduled tasks in alpha, alongside chat on files and tools.[1][2]

Claude Cowork migration path An official guide with agent-run scan, plan, and apply steps moves Cowork plugins, SKILL.md files, and connectors; history, credentials, and non-GitHub plugins stay behind.[1]

Self-hosting or managed cloud The platform self-hosts with Helm or Docker Compose and also runs as the browser-based Cloud app, with the vendor charging the same Team price either way.[1][2]

Team controls on paid plans Team adds SSO/SAML, an extension marketplace, and cloud automations; Enterprise adds SCIM, audit logging, desktop policies, white-labeling, and bring-your-own inference.[1]

If your work already spans Claude, GPT, Gemini, and local models, one gateway with per-team keys and limits directly serves that shape; bringing a single hosted key alone does not provide the same boundary.[1][2]

The MIT desktop app and the first five Cloud seats cost nothing, so a team of up to five can adopt the stack without procurement; larger production use still needs a subscription.[1]

Centrally managed skills, connectors, desktop policies, and version controls address the sprawl of everyone wiring their own tools; these are documented options, not evidence from our own deployment.[1][2]

Production use of the ee/ control plane beyond five users, development, or the 30-day evaluation requires a subscription under the EE License, so the MIT label does not cover the whole suite.[1][2]

Conversation history, saved credentials, plugins outside public GitHub repositories, and existing scheduled tasks do not transfer and must be recreated by hand; Dispatch and live artifacts are listed as coming soon.[1]

Model costs stay with the adopter through bring-your-own keys unless the $10-per-user managed inference add-on is bought, and hosted inference sends context to the selected provider.[1][2]