Local desktop use and team sharing have different terms
OpenWork's desktop app can work without an OpenWork account, while its hosted MCP connection asks you to sign in and choose an organization. Decide which route you need before treating OpenWork as one deployment. Someone trying an agent on local files has a different evaluation ahead of them from an administrator distributing capabilities across colleagues' tools.
The README describes a desktop workspace built on OpenCode, with provider connections and local models through Ollama. That makes it interesting if you want to choose the model while keeping a dedicated place for file-based work. We did not measure whether an agent completed a task correctly or how well any model followed instructions.
Shared capabilities can reach the agent you already use
The documented OpenWork MCP exposes capability search and execution, plus tools to list skills and read a selected SKILL.md. Its examples cover Codex, Claude Code and OpenCode. A team can therefore investigate distributing skills without requiring every colleague to adopt the desktop interface.
The practical question is what each person can discover and execute after signing in. Den documents assignments to organizations, teams and individuals, along with provider access and desktop policies. For an evaluation, choose one intended capability, assign it to a limited audience and check its behavior from the client those people use. This is a proposed trial, not an action completed in our sandbox.
The README's local-file language also deserves a careful reading. It supports local models as an option, alongside remote providers and connected services. A deployment decision still needs the chosen model's data handling and each connection's permissions. Our evidence contains no network inspection or security assessment.
What happened when we ran it
Our October 1, 2026 run at commit 53f090c completed dependency installation in 159 seconds, recording 1,769 dependencies and 2,364 MB on disk. The source build exited 1 after 188 seconds, and the test command exited 1 after 7 seconds. Those are repository setup results, with no recorded desktop task or authenticated capability execution.
The test log reached the app's Bun-based command and printed sh: 1: bun: not found, followed by a spawn error and the failed lifecycle status. That establishes a missing executable at that stage. It provides no passing assertion count and no basis for saying the product's application tests exposed incorrect behavior.
The supplied build tail contains a child-process failure, JavaScript stack frames and the lifecycle exit. It does not identify the command's underlying error. Assigning the build failure to Bun as well would go beyond this evidence. Getting dependencies onto disk was successful; completing the source build was not.
The configured sandbox had 3 CPUs and 8 GB RAM under image label lab-node:22. The actual build tail says Node.js v24.19.0. The image name therefore cannot establish the runtime version. The checkout scan recorded 42 CI workflow files and monorepo workspaces. Its failure to detect a tests directory does not establish an absence of coverage: the current README locates executable coverage under evals/specs.
Den changes the licensing decision
The README places code outside ee/ under MIT and OpenWork Den under the OpenWork EE License. Den's code is source-available, with production subscription requirements and exceptions. The stated exceptions include organizations with up to five users, evaluation for 30 days at any size, and development use. Enterprise releases convert to MIT after two years.
If your procurement rule requires an MIT team control plane, that distinction can end the evaluation early. If your organization qualifies for an exception, read the applicable terms before assuming the exception covers the deployment you intend. The free local desktop does not settle Den's licensing for a larger organization.
Current releases do not validate our older checkout
GitHub records a November 3, 2026 push, and the recent activity feed includes pull requests updated that day. The October 26 v0.18.63 release describes Claude-history imports and connected-file transfers requiring approval. Together these show ongoing development, rather than support for an abandonment claim. They do not establish response times or released-product reliability.
That release came after our October 1 source run. Its transfer and import features remain upstream descriptions here. If those features are why you want OpenWork, evaluate the intended release with representative files and permissions instead of using our source timings as a proxy for their behavior.
OpenCode and Ollama answer narrower requirements
OpenCode describes an open-source coding agent and offers terminal and desktop interfaces. Start there if coding is your main requirement and OpenWork's organization assignments add little to your workflow. OpenWork itself uses OpenCode, so this is a choice about the surrounding workspace and sharing features.
Ollama documents a REST API for running and managing models. Choose it when the missing component is model access for an application you already have. It is also a documented local-model option for OpenWork. We have no comparative speed or reliability measurements for either alternative.
Documentation earns 3/5 on a limited editorial scale: 1 means a route cannot be followed from the material read, 3 means useful instructions with an unresolved requirement, and 5 requires instructions verified across the reviewed routes. OpenWork explains its routes well, but our observed Bun requirement remains unresolved by the contributor prerequisite list. Other dimensions remain unassessed. Before rollout, make the proposed file task or shared capability succeed with the permissions your colleagues will actually have.

