Use OpenWorkBuddy to deliver work you can inspect.

1. Install and open the app

Download the latest installer from the official GitHub Releases page. Choose the package for macOS or Windows. On Linux, run from source or use Docker. OpenWorkBuddy can also run as a self-hosted server for a team.

1. Install and launch OpenWorkBuddy.
2. Open Settings and configure a model provider or local model.
3. Choose a data location with enough disk space and include it in your backup plan.

Operating systems may show a first-run security prompt because community builds are not signed with a paid vendor certificate. Follow the release-specific instructions in Install and launch.

2. Connect a model

In Settings → Models, add a provider and model, enter its API key, then choose Test connection. Use a hosted API, a compatible gateway or a local model such as Ollama.

• Use the provider's exact model ID and API base URL; presets fill common protocol fields.
• Pick a model with enough context for the documents and instructions you provide.
• Cloud requests are processed under the selected provider's terms. Local-first describes where the workspace runs, not where every model runs.
• Do not put credentials in task prompts or public issue reports. Store model keys in app settings and keep backups private.

See Model configuration for protocol formats, local models and troubleshooting.

3. Describe a task clearly

State the deliverable, audience, source material, constraints and output location. Say what a finished result must contain and how you will review it.

Example: “Read the attached quarterly reports and make a six-slide leadership review. Use figures from the source files, flag missing values, save it in this project, and check the actual deck before finishing. Do not invent metrics.”

Attach source files or add reference folders to the current project's context. Keep different clients or workstreams in separate projects.

If an action needs your decision, review the approval card and its displayed changes. Press Esc to reject; the primary modifier plus Enter allows an action once. Review important results even when the Agent reports success.

In Goal mode, the execution plan appears inside the conversation as a collapsed list. Click its header to expand the acceptance criteria in place. Progress updates keep your expansion and keyboard focus, and the list scrolls with the conversation. If execution pauses, its reason and resume button remain visible.

4. Review delivered files

Use the results panel to preview or locate files in the workspace. Check facts, formatting, formulas, links and media in the actual file. Ask for a targeted revision and keep an earlier version when you need a rollback.

History is grouped by date. Pin important tasks and files in the sidebar. The inbox gathers approvals, questions, errors, running work and unread tasks; open a scheduled run to inspect its execution history.

Use ⌘N / Ctrl+N for a new task and ⌘K / Ctrl+K to search history.

5. Organize projects and reference files

Create a project for a client, product or recurring workflow. Add project instructions, source folders and library references that should carry across its tasks. Pin frequently used files for quick access.

The library stores reference material separately from generated work. Check a file's license and confidentiality rules before importing it. On a shared server, confirm your organization's role and sharing settings before adding shared material.

See Multi-user collaboration for account roles, workspaces, quotas and shared settings.

6. Use schedules and the inbox

Create a one-time or recurring task from the Automations page. Give it a clear name, an unambiguous time and a description that can run without extra chat context.

• A schedule created from a conversation writes its result back to that conversation.
• A standalone schedule puts its result in the inbox; open it there to see run history.
• Before enabling an unattended task that changes files or contacts a service, check its instructions and last run.

See IM and scheduled tasks for notification channels.

7. Protect your data

OpenWorkBuddy can read files, run commands and use connected services. Grant only the folders and connectors a task needs. Inspect changes shown on approval cards and review third-party Skills or plugins before installing them.

• Back up the OpenWorkBuddy data folder and important workspace files.
• Keep API keys, access tokens, session exports and private customer data out of public issues and screenshots.
• Share only the material needed when using cloud models or remote connectors.
• For an internet-facing server, use HTTPS, strong authentication and the deployment security guidance.

Read the full Security guide and Remote access guide.

Connectors and OAuth

1. Open Settings → Connectors. Choose a preset, paste MCP JSON, or import a selected local configuration on a single-user desktop.
2. For Streamable HTTP connectors, add the endpoint and click OAuth authorization. Confirm the requested scopes on the provider page, then return to the workbench. Some providers require a registered Client ID; expand the client fields if necessary.
3. The workbench uses PKCE, checks the authorization server, stores tokens in a private local file and refreshes expired access tokens. Removing local authorization stops its use here; revoke the provider-side grant from your account settings there.
4. Each connector can be disabled, reconnected or edited. Set each tool to allow, ask or deny. An unavailable tool is removed from the model's offered tools.

Native OAuth applies to HTTP connectors used by the built-in engine. Existing command-based OAuth helpers continue to manage their own authorization. A local CLI engine may not support a remote connector; inspect its connector notice before running.

Region annotations and file changes

1. Open a delivered document, table, image or webpage in the preview. Click Annotate a region and drag over the area to change.
2. Write the requested change and choose Add to change request. The draft includes the file, rectangle, scroll position and available text, table cells or webpage selectors. Review and send the draft.
3. The request instructs the Agent to preserve other content and ask when the target is unclear. Visual coordinates can become stale after a layout change; review the revised file. PDF and image previews provide coordinates rather than text selectors.
4. Under a file-changing step, open Review changes. Click Revert this block to restore that change only; Undo this block revert brings it back. Binary files and files over the text limits use the full step rewind.
5. Click a line to add a line-specific revision request. If a task is running or the file has independent edits, block reverts are refused to protect the current file.

Team tasks and read-only sharing

In organization mode, choose Team before sending the first message of a new task. Team tasks use shared organization API models and administrator-managed connectors through the built-in engine. Personal subscription engines and personal model keys are excluded. Once a task has records, its billing scope stays fixed.

Costs opens an inline monthly summary of your team and personal tasks. Model calls and paid tools carry the scope and task ID in their usage records. Unknown prices and unconnected personal subscription or own-key bills are counted separately; these are configured-price estimates, and the provider bill remains authoritative. Existing organization and member limits still apply.

The task owner can create a read-only link from the history menu. Only signed-in members of the same organization can view the conversation and delivered files. Tool arguments and configured credentials are excluded. Revoke the link in the inline sharing panel to stop access. Single-user mode hides this feature.

Record a reusable browser skill

1. Install the included Chrome extension from Settings → Browser, then pair it with the workbench.
2. Open the page in Chrome. In Settings → Browser, click Record current tab. The extension shows REC while recording. Perform the workflow yourself, then return and stop.
3. Review the generated skill. It records click locations, scrolling and field locations, with input values replaced by runtime parameters. Password fields, URL parameters, cookies and page text are not recorded. Recording stops on another site, when disconnected, or at the limits.
4. Fill in a useful name, revise the steps, then save the skill or download its Markdown draft. Saving a shared skill requires the platform administrator role. Each later run still observes browser action approval rules.

Only the site's origin is retained for navigation. Supply the starting page when running the skill. Page structures can change; recorded locators are hints that must be checked against freshly read elements.

12. Troubleshoot and get help

• For model connection errors, verify the base URL, protocol, model ID and key, then test again.
• If an output is missing, check the active project and workspace before rerunning.
• If a task is waiting, open the inbox and review its question or approval.
• For a reproducible bug, open a GitHub issue with the version, operating system, steps and sanitized logs. Remove keys, personal data and file contents first.