OpenWorker Review

OpenWorker is a desktop AI agent that you ask for an outcome rather than an answer: a customer brief from your CRM and inbox, a report, a rescheduled calendar, a Slack thread reply, or a security review of a repository with a fix branch ready. It breaks the task into steps, works across your local files, terminal and connected apps, and returns the result in the app, in Slack, or as a file (Markdown, PDF, documents, spreadsheets or web pages). Before anything consequential, such as sending a message, changing a calendar or running a command, it stops for a check-in you approve, and you can set how independently it may act. The project ships specialist Security coworkers first (Security Reviewer, Cloud Posture and Dependency Audit), alongside general coworkers for sales, executive assistant, marketing and on-call work, and you can describe your own.

Setup is a download: a signed macOS build for Apple Silicon (macOS 12+) and a Windows 10/11 x64 build that is not yet code-signed, so SmartScreen warns. Open the app, add an API key for a model provider or point it at Ollama, and connect tools with one-click or manual connections. The agent loop, conversations, model keys and connector tokens stay on your machine; an optional free OpenWorker account only brokers OAuth sign-ins for connectors such as Slack. The source is on GitHub under the MIT license, the engine is built on the aisuite Python library, and the README labels the app an open beta.

OpenWorker has no paid plan. The desktop app is open source and free; you pay your model provider directly, or nothing if you run a local model.

Model usage is billed by whichever provider you choose (for example OpenAI, Anthropic or Google), under that provider's own pricing and terms; OpenWorker's terms say it is not a party to that relationship. OpenRouter can be used either with a pasted API key or by signing in with an OpenRouter account and spending its credits. Running a local model through Ollama, llama.cpp, vLLM, LM Studio or similar avoids per-token charges. The FAQ answers "Is it really free?" with yes and says you pay "nothing to us".

Prices checked on the vendor's pricing page on November 10, 2026. OpenWorker has no separate pricing page; see the FAQ on openworker.com and the Terms of Service.

What OpenWorker's site and README describe:

OpenWorker suits people who want an agent to finish knowledge-work tasks across their own tools while keeping data and keys on their computer: engineers and security teams who want scanner-backed repo and Terraform reviews delivered as pull requests, and operators, founders or sales staff who want call prep, reports, inbox and calendar work done with approval before anything goes out. It also fits anyone who wants to use cheaper or local models instead of a fixed subscription, and developers who want to read a production-shaped agent harness.

If you want a hosted agent that runs in the cloud without installing anything, Manus is the managed option. If you mainly want an open-source agent for coding and engineering tasks, Goose or Open Interpreter are closer fits. For no-code business agents with hosted integrations, look at Lindy. Teams that need signed Windows builds or a vendor support contract should note OpenWorker is still in open beta.

Coverage so far is launch write-ups rather than long-term user reviews. A DEV Community explainer (July 29, 2026) praises the typed permission design, explicit prompt-injection handling and the local-first, bring-your-own-key setup, but stresses that the project is early ("Expect rough edges"), that Windows builds are unsigned, and that it targets knowledge work rather than coding. A second DEV Community post by an independent consultant (July 2026, edited September 2026) calls OpenWorker a security improvement over OpenClaw, citing risk-classed tool calls and approval-gated MCP servers, while flagging that connector tokens were then stored in a plaintext file, that bypass mode gives full shell access, and that the app is in beta; OpenWorker's README now says tokens live in the app's local secret store. No independent benchmark of task success rates was found.