OpenWorker is a desktop agent application for macOS and Windows that carries a task through to a deliverable: a reviewed codebase with proposed fixes, a document or spreadsheet, a Slack reply, or a triaged inbox. You describe the outcome, the agent breaks it into steps, works across local files, the terminal, and connected apps, and checks in before any consequential action such as sending a message, changing a calendar, or running a command.
The app is local-first. The agent loop, conversations, connector tokens, and model keys stay on the machine in a local secret store; the only hosted component brokers OAuth handshakes for connectors, and connectors can also be configured with manually created credentials. Model access is bring-your-own: OpenAI, Anthropic, Google Gemini, DeepSeek, Qwen, Mistral, Grok, Kimi, MiniMax, GLM, and others, open-weight models via Together and Fireworks, or fully local models through Ollama.
Governance is built into the architecture in three tiers. A set of dangerous or irreversible operations is always human-only, regardless of mode. Other actions are approval-gated by default, and one-off approvals can graduate into standing rules and then config allowlists; in auto-approve mode a reviewer model passes routine actions and escalates uncertain ones, with a circuit breaker that pauses the reviewer after repeated denials. Every tool call is recorded with its approval provenance and the reviewer reasoning, and unattended runs park their requests in an inbox rather than self-approving.
Under the hood, a native shell and GUI (React with a Tauri shell) supervise a local Python agent server built on the aisuite library, which provides a unified chat-completions API and an agents layer with tools and MCP support. Specialist security coworkers ship first: security review combines deterministic scanners such as semgrep with model reasoning and re-scans proposed fixes before approval; cloud posture and incident triage are further presets.
Features
- Finished deliverables: documents, spreadsheets, reports, and web pages land as files to open and share
- Slack entry point: mentioning @OpenWorker in a channel opens a desktop session and returns the answer as a thread reply
- Connectors: 25+ integrations including GitHub, Slack, Jira, Notion, Linear, HubSpot, Outlook, monday.com, Gmail, and Google Calendar, plus any MCP server with per-tool control
- Scheduled automations: morning briefs, weekly reports, or a standing watch on a channel, with full transcripts
- Approval gating: writes, sends, and shell commands wait for approval, with an optional auto-approve mode that still escalates
- Hard floors: irreversible operations always escalate to a person
- Audit trail: each tool call is stored with who approved it and why
- Bring your own model: commercial providers, open-weight hosts, or local Ollama models, with a curated list of models verified for tool calling
- Security coworkers: codebase and dependency review, cloud configuration audit, and incident triage presets
- Voice input: a Rust speech-to-text sidecar
