OpenWork
Give AI agents a desktop workspace for working on local files and sharing capabilities with a team.
- Source repo
- different-ai/openwork
- Stars
- ★ 24k
- Last updated
- today
- License
- NOASSERTION
- Primary language
- TypeScript
- FA score
- 71/100 · Some gaps
At a glance
- How it runs
- Works with
- Universal · cross-platformCodex · Claude Code
- Cost
- Free tier plus a paid hosted plan
- Setup effort
- Medium · a few setup steps
- You'll need
- Typical use
- Individuals who want an agent to work on local files in a macOS, Windows, or Linux desktop workspace.
- Not a fit if
- Users who need a CLI-only interface
- Teams that require the organization control plane to be MIT-licensed
- Source review
- 71/100 · Some gaps
What does this agent do, and when should you use it?
OpenWork is a desktop app for macOS, Windows, and Linux, built on OpenCode, that lets AI agents work on a user's own files. It supports more than 50 model providers, with options including your own API keys, ChatGPT sign-in, and local models through Ollama or OpenAI-compatible servers. Its OpenWork MCP also makes skills, plugins, MCP connections, and some connected services available from compatible agents such as Codex, Claude Code, and Cursor. Teams can use OpenWork Den to manage members, model access, desktop policies, and shared capabilities, and can self-host the control plane. The desktop app and core platform are MIT-licensed; OpenWork Den under ee/ uses the OpenWork EE License, with separate terms for production use.
After installing OpenWork, a user creates a workspace and selects a model provider and model so an agent can work on local files. Administrators can use OpenWork Den to invite members, create teams, configure model access and desktop policies, and publish skills and plugins. After configuring the OpenWork MCP endpoint at https://api.openworklabs.com/mcp/agent in a compatible client, users can call search_capabilities to find available capabilities, execute_capability to run one, list_skills to list skills, and get_skill to read a SKILL.md. The repository also contains an Electron desktop app, a React interface, and openwork-server; contributors can launch the development app with pnpm dev.
- Individuals who want an agent to work on local files in a macOS, Windows, or Linux desktop workspace.
- Model users who want to bring API keys, sign in with ChatGPT, or use a local model through Ollama.
- Codex or Claude Code users who want to access assigned skills and MCP connections from those clients.
- Organization administrators who need to assign skills, plugins, and model-provider access to members and teams.
- Organizations that want to run the team control plane on their own infrastructure.
How do you install or deploy this agent?
Install the desktop app from the OpenWork download page; the README also links to GitHub releases. For a source development build, the repository documents these steps:
How do you use this agent?
Open OpenWork, create a workspace, and connect a model provider. Alternatively, add the remote MCP server to Codex or Claude Code. The client then opens a browser so you can sign in and choose an OpenWork organization.
What are this agent's strengths and limitations?
- Offers desktop apps for macOS, Windows, and Linux, with agents working on the user's own files.
- Supports more than 50 model providers, bring-your-own API keys, ChatGPT sign-in, and local models through Ollama or OpenAI-compatible servers.
- The OpenWork MCP exposes skills and MCP connections to compatible agents such as Codex and Claude Code.
- OpenWork Den offers team administration and a self-hosting option for the control plane.
- The desktop app is not a CLI, so users who only want a terminal interface may not find it suitable.
- Using a model requires a provider access method, such as an API key, ChatGPT sign-in, or a local model server.
- OpenWork Den under ee/ uses the OpenWork EE License; production use generally requires a subscription, with exceptions such as free use for organizations of up to five users.
- The README does not give a step-by-step first-run setup for the downloaded desktop app.
How does this agent compare with similar options?
The README presents OpenWork as an open-source alternative to Claude Cowork and Codex, and says Claude Cowork skills, compatible plugins, and MCP servers can be migrated. It explicitly distinguishes OpenWork from Claude Cowork by offering desktop apps for macOS, Windows, and Linux, support for multiple model providers, and a self-hostable team control plane.
Key facts side by side with the most closely related agents.
| Agent | Source review | Form / cost | Stars | Updated | Language | Full support on |
|---|---|---|---|---|---|---|
| OpenWork This agent | 71 · Some gaps | Desktop appFreemium | ★ 24k | today | TypeScript | Codex · Claude Code |
| PocketPaw | 46 · Major gaps | Desktop appFree + model costs | ★ 892 | 1mo ago | Python | Codex · OpenAI API · Claude API |
| Conduit | 73 · Some gaps | Desktop appFree + model costs | ★ 2.3k | 4d ago | Dart | OpenAI API |
| NextClaw | 58 · Major gaps | Desktop appFree + model costs | ★ 262 | today | TypeScript | Codex · Claude Code |
How does FollowAgents rate this agent?
Why each dimension lost points
The docs say local files can remain local with cloud optional, Den-managed MCP credentials are encrypted and kept in Den, and authorization discovery is side-effect free and does not silently widen scopes. Install and development guidance also describes token isolation and loopback-restricted access. Deductions: these controls are mostly documented assertions; the supplied material does not show core permission implementations, complete data flows, or dependency audits, and confirmation and rollback are explicit only for some flows.
The repository specifies Node/pnpm versions, CI contract checks, an evaluation suite, and concrete troubleshooting steps, with several failure-handling paths described. Deductions: the supplied files cover only part of the product and cannot establish dependency availability or consistency across all paths; no execution results are provided.
The README identifies desktop support for macOS, Windows, and Linux, multiple model and credential options including local models, team and self-hosted use, and access from other agents. Docs address developer, administrator, and MCP-client scenarios. Deductions: capability boundaries and use limits appear in places but are not a complete product boundary map, and trigger conditions are mostly summarized.
The repository provides a layout map, contributor guidance, install steps, operator docs, troubleshooting, and directory-split licensing; it also describes release channels and version handling. Deductions: the supplied material lacks a clear release history or change summary, maintenance ownership appears mainly as a security contact and limited document-owner labels, and readers must reconcile top-level license metadata with directory-specific terms.
The product combines a desktop workspace, cross-tool MCP, shared team capabilities, and an administration plane; the docs include install and connection examples that show practical use. Deductions: benefits and costs are not quantified, value claims are mostly project-authored, and the evidence is insufficient to assess actual user outcomes or net benefit.
Several security, installation, and protocol claims have corresponding details across docs, workflows, and test specifications; some docs state test scope and explicitly identify what is not certified. Deductions: evidence remains largely self-reported within one repository, supplied test files do not prove execution, and broad product claims lack independent or cross-source confirmation.
- The README says files stay local, while the materials also describe cloud inference, Den, and remote MCP; confirm which data is sent to external services for the specific deployment and workflow.
- Licensing is split by directory: ee/ uses the OpenWork EE License and other content is MIT. Do not infer an individual component's license from repository-level NOASSERTION or the README summary alone.
- This assessment uses only the supplied static materials and did not run code or verify release artifacts; security and feature claims are not independently validated findings.