Open Computer Use (MCP desktop service)
An MCP service for running desktop Computer Use operations from AI clients.
- Source repo
- iFurySt/open-computer-use
- Stars
- ★ 2.9k
- Last updated
- 3d ago
- License
- MIT
- Primary language
- Swift
- FA score
- 44/100 · Major gaps
At a glance
- How it runs
- Works with
- Universal · cross-platformCodex · Claude Code
- Cost
- Free software; you pay for model usage
- Setup effort
- Medium · a few setup steps
- You'll need
- Typical use
- A Codex App or Codex CLI user who wants to connect a local Computer Use service through MCP.
- Not a fit if
- Users on macOS 13 or earlier
- Users unwilling to grant Accessibility and Screen Recording permissions
- Source review
- 44/100 · Major gaps
What does this agent do, and when should you use it?
Open Computer Use is an open-source Computer Use service packaged as an MCP server. It is documented to let AI agents and MCP clients run Computer Use on macOS, Linux, and Windows. The npm package provides the open-computer-use command and the ocu alias, plus installers for Codex, Claude Code, Gemini CLI, and opencode configuration. Its CLI can invoke one tool or execute a sequence in one process so element_index state can be reused, printing MCP-style JSON results. The macOS runtime requires macOS 14.0 or later and requires Accessibility and Screen Recording permissions on first use.
The service is started with open-computer-use or ocu and exposed to clients with the open-computer-use mcp command. Documented calls include list_apps, get_app_state with an app argument such as TextEdit, and press_key with app and key arguments. open-computer-use call executes one Computer Use tool and prints an MCP-style JSON result. The --calls and --calls-file options execute multiple tool calls in one process, allowing element_index state to be reused. Built-in install commands write MCP configuration for Codex, Claude Code, Gemini CLI, or opencode.
- A Codex App or Codex CLI user who wants to connect a local Computer Use service through MCP.
- A Claude Code user who needs install-claude-mcp to add the service to ~/.claude.json.
- A Gemini CLI user working in a project who wants to write the integration to ./.gemini/settings.json.
- An opencode user who wants the MCP server added to ~/.config/opencode/opencode.json or the active configuration file.
- A desktop automation user who needs to inspect TextEdit and then send Return in one reusable-state call sequence.
How do you install or deploy this agent?
Install globally: npm i -g open-computer-use. Start it with open-computer-use or ocu. On macOS, the runtime requires macOS 14.0 or later; run it once and grant Accessibility and Screen Recording permissions. For Codex, run open-computer-use install-codex-mcp. Manual MCP configuration is: {"mcpServers":{"open-computer-use":{"command":"open-computer-use","args":["mcp"]}}}.
How do you use this agent?
Start open-computer-use, then connect an MCP client to command open-computer-use with args ["mcp"]. Test a tool call with open-computer-use call list_apps. Query an application with open-computer-use call get_app_state --args '{"app":"TextEdit"}'. Run a sequence with open-computer-use call --calls '[{"tool":"get_app_state","args":{"app":"TextEdit"}},{"tool":"press_key","args":{"app":"TextEdit","key":"Return"}}]'.
What are this agent's strengths and limitations?
- It exposes Computer Use through MCP, and the documentation states that any AI agent or MCP client can use it.
- It explicitly documents macOS, Linux, and Windows support and includes demonstrations for the platforms.
- Built-in installers target Codex, Claude Code, Gemini CLI, and opencode rather than requiring every integration to be written by hand.
- Sequence execution through --calls or --calls-file preserves element_index state within one process.
- Single CLI calls print MCP-style JSON, which is useful for inspecting scripted tool invocations.
- The macOS runtime is limited to macOS 14.0 or later and requires Accessibility plus Screen Recording permissions on first use.
- Client setup changes user or project configuration files, including ~/.codex/config.toml, ~/.claude.json, and ./.gemini/settings.json.
- The supplied material shows only a few callable tool examples and does not provide a complete tool catalog or failure-result schema.
- The supplied material does not detail permissions, desktop-environment assumptions, or dependencies for Windows and Linux.
- No hosted offering, SLA, or commercial support is described in the supplied material.
How does this agent compare with similar options?
The project says it was inspired by OpenAI Codex Computer Use and demonstrates use in Codex App and Codex CLI, describing the experience as matching the official one. Its documented delivery model is a locally installable open-source MCP service; the supplied material does not provide a detailed feature-by-feature comparison.
Key facts side by side with the most closely related agents.
| Agent | Source review | Form / cost | Stars | Updated | Language | Full support on |
|---|---|---|---|---|---|---|
| Open Computer Use (MCP desktop service) This agent | 44 · Major gaps | CLIFree + model costs | ★ 2.9k | 3d ago | Swift | Codex · Claude Code |
| Clawd Cursor | 74 · Some gaps | MCP serverFree + model costs | ★ 406 | 1d ago | TypeScript | Codex · Claude Code |
| macOS Harness | 55 · Major gaps | CLIFree + model costs | ★ 1k | 22d ago | Python | Codex · Claude Code |
| SwiftAutoGUI Desktop Automation | 60 · Some gaps | CLIFree + model costs | ★ 94 | 19d ago | Swift | Claude Code · OpenAI API |
How does FollowAgents rate this agent?
Why each dimension lost points
Evidence: README mentions macOS Accessibility and Screen Recording permissions, but does not explain least privilege or permission scope. There is a doctor command to check permissions, but no explicit user confirmation mechanism. Data flow transparency is lacking; no explanation of how data is processed or transmitted. Sensitive data handling is not mentioned. Dependency security is not mentioned, but release workflow uses pinned action SHAs, showing some security awareness. External effects: the tool controls the system, but no restrictions are stated. Rollback mechanism is not mentioned. Source attribution: README mentions inspiration from OpenAI, but third-party code attribution is unclear. Deductions: lack of clear evidence for least privilege, user confirmation, data flow transparency, sensitive data handling, dependency security, external effects control, rollback, and source attribution.
Evidence: Commands in README and package.json are consistent, such as open-computer-use and ocu alias, but no detailed error handling or failure messages are provided. Dependency availability: npm package and GitHub Releases exist, but not verified. Failure messages are not mentioned. Deductions: self-consistency is adequate, but dependency availability and failure messages lack evidence.
Evidence: README clearly targets AI agents and MCP clients, supports macOS, Linux, Windows, and provides multiple installation methods. Capability boundaries are not explicit, but tool list is provided. Trigger precision: call command and argument examples are given. Environment fit: multi-platform support, but specific system requirements are not detailed. Deductions: capability boundaries and trigger precision lack evidence.
Evidence: README is well-structured with quick start, more commands, demos. Installation notes are detailed, including npm global install and MCP config. Naming stability: CLI alias provided, but version naming not explained. Examples and FAQ: multiple demos and command examples, but no FAQ. Known limitations not mentioned. License is MIT with LICENSE file. Versioning/changelog not mentioned. Maintenance responsibility: SECURITY.md exists but is a template. Deductions: known limitations, versioning/changelog, and maintenance responsibility lack evidence.
Evidence: Output usability: MCP tools and CLI output JSON results. Marginal value: as open-source alternative, provides similar functionality. Cost-benefit: no performance or resource consumption information. Deductions: cost-benefit lacks evidence.
Evidence: README claims are supported by demo videos and screenshots, but no verifiable test results. Cross-source corroboration: star history and badges, but no independent verification. Fact-inference separation: README distinguishes facts and inspiration, but not explicit. Deductions: claim traceability, cross-source corroboration, and fact-inference separation lack evidence.
- This tool requires macOS Accessibility and Screen Recording permissions, which may pose security risks; grant with caution.
- No data flow transparency or sensitive data handling details are provided; assess data privacy risks before use.
- No rollback mechanism is provided; operations may be irreversible.
- Dependency security is not explicit; check for dependency vulnerabilities.
Recent releases
Latest entries from the project's GitHub Releases. Notes are quoted from upstream as written, not rewritten or tested by FollowAgents.
-
Release notes excerpt
What's Changed * Improve macOS automation with more accurate coordinate clicks, focused-sheet snapshots, live-focus text replacement, faster accessibility reads, and more reliable permission guidance. * Add macOS text selection and optional accessibility-tree or screenshot collection across platforms; improve Windows scrolling and scaled-display input, and reduce Linux observation overhead. Virtual displays now require explicit opt-in. * Preserve notarization tickets for installable Locked Use components and refresh packaged artifacts after notarization. …View on GitHub ↗ -
Release notes excerpt
What's Changed * Kept the macOS software cursor aligned with moving windows during motion, click feedback, and idle waits. * Fixed Linux action read-back configuration so batch operations can skip automatic snapshots while preserving the last observed element indexes. * Updated package and installation links to the renamed open-computer-use repository while preserving existing package names, CLI aliases, and app identity. Full Changelog: https://github.com/iFurySt/open-computer-use/compare/v1.1.1...v1.1.2View on GitHub ↗ -
Release notes excerpt
What's Changed * Improved macOS Locked Use authorization and installation, with a plug-in load check before activation and clearer component version diagnostics. * Reduced idle polling and repeated signature checks, and fixed a socket cleanup issue that could disconnect the automation client. * Fixed removal and replacement of older Locked Use components while preserving native password authentication. Full Changelog: https://github.com/iFurySt/open-codex-computer-use/compare/v1.1.0...v1.1.1View on GitHub ↗