OpenWorker Functional and Technology Baseline.

Baseline Version: OpenWorker v0.1.6 (current official Beta trunk)
License: MIT Open Source, permits commercial use and derivative development.
Positioning: Local‑First desktop‑based human‑AI collaborative AI Coworker (local intelligent agent for knowledge workers).
Purpose: architecture review, project technology selection, HermCoze system integration assessment, and development requirement boundary definition.

I. Overall Baseline Statement

This document defines the stable, production‑ready capabilities of the official release. Experimental features not yet merged into the main branch are excluded from the baseline and shall not be considered part of the production roadmap.

Platform Compatibility Baseline
macOS (Apple Silicon): full feature set, code‑signed, automatic updates.
Windows x64: feature‑aligned with macOS, but lacks digital signing – may trigger security warnings and system interception prompts.
Linux: source‑based server‑side execution only; no officially packaged GUI client.

Core Design Invariants – the following four architectural tenets are firmly upheld and will not be refactored in the near term:Local‑First,Model‑Neutral,Human‑in‑the‑Loop ,Deliverable‑Oriented (finished work, not just text).

II. Technology Stack Baseline (Fixed Component Versions and Technology Selections)

2.1 Overall Layered Technology Stack
Tauri GUI Frontend Layer
                         ↓
Local API Gateway & Permission Control Service (Python)
                         ↓
TurnEngine Agent Core Runtime
                         ↓
aisuite LLM Unified Abstraction Layer
                         ↓
MCP Model Context Protocol Connector Layer

2.2 Frontend Desktop Layer
• Tauri v1.x (Electron alternative – lightweight binary packaging)
• UI Framework: React + TypeScript
• Communication: Tauri IPC for local inter‑process communication
• Capabilities: task visualisation, approval pop‑ups, system notifications, file invocation, and task flow log display

2.3 Backend Core Service Stack (Python Foundation)
• Python 3.11 ~ 3.12 (baseline locked range; Python 3.10 and below are not supported)
• Asynchronous Framework: native asyncio task scheduling
• Local Persistence: SQLite (task snapshots, session logs, approval records)
• Secure Credential Storage: leverages OS keychain (Keychain on macOS / Windows Credential Manager)
• Web Service: lightweight built‑in uvicorn local loopback service – accessible only from localhost, no public port exposure

2.4 Agent Core: TurnEngine
• Proprietary task‑loop engine, with no mandatory external dependencies on similar frameworks
• Built‑in capabilities: task planning, iterative reflection, execution state snapshots, and HITL approval interceptor

2.5 LLM Abstraction Layer: aisuite (Official Project Dependency)
• Unified encapsulation of various LLM APIs and standardised Function Calling
• Baseline compatibility: OpenAI‑compatible protocols, Anthropic, Google Gemini, and local models via Ollama
• Built‑in features: request retries, rate limiting, error handling, and hot‑swappable model configuration

2.6 Tool Communication Standard: MCP (Model Context Protocol)
• Adopts the standardised MCP protocol as the unified communication specification for all connectors
• All built‑in connectors and third‑party extensions comply with the same protocol, enabling hot‑plug integration

2.7 External Dependency Boundaries (Baseline Constraints)
•  Optional integration: Ollama (local offline large language models)
•  Optional extension: external standalone MCP services (e.g., MCP‑RAG, custom business connectors)
•  Not included in baseline: vector databases, Redis, message queues, web service gateways, or multi‑tenant management modules

2.8 Packaging and Deployment Baseline
• Delivery format: desktop client installer package
• Runtime mode: GUI client runs in foreground; can be minimised to system tray (scheduled tasks rely on the process remaining alive)
• Not supported: container image deployment, or headless server‑style persistent operation (headless mode is for debugging purposes only)

III. Functional Baseline (Stable Capabilities in v0.1.6)

3.1 Foundational Platform Capabilities Baseline
• BYOM – Model‑Neutral
◦ Cloud models: OpenAI, Claude, Gemini, DeepSeek, Kimi, GLM, xAI Grok, Mistral, TogetherAI.
◦ Local offline: native integration with Ollama, supporting fully offline operation without internet connectivity.
◦ Constraint: a single task supports only one model at a time; parallel multi‑model inference is not supported.
• Local‑First Data Policy
◦ Sessions, task snapshots, and credentials are encrypted and stored locally on the machine.
◦ User original files are never automatically uploaded to third parties; only necessary payloads are transmitted on‑demand when invoking LLMs or external SaaS services.
◦ Task checkpointing supports resumption after application restart, allowing incomplete tasks to be recovered.
• Comprehensive Audit Logging
◦ Records instructions, all tool invocations, input/output parameters, and human approval actions. Logs are exportable for external review.

3.2 TurnEngine Agent Core Capabilities
• Goal‑Driven Automatic Task Decomposition
◦ Users input a natural‑language final objective; the engine automatically decomposes it into a sequence of sub‑tasks. The default maximum iteration count is 12 tool‑call rounds (configurable). No manual step‑by‑step prompting is required.
• HITL (Human‑in‑the‑Loop) Permission System – Core Differentiator
◦ Five operation modes: discuss / plan / interactive (default) / auto / custom.
◦ High‑risk action interception rules are hard‑coded: actions such as sending external emails, IM messages, executing Shell commands, or modifying online SaaS resources trigger a mandatory pop‑up for human confirmation. The security checks cannot be fully disabled.
• Session‑Level Short‑Term Task Memory
◦ Constraint: no cross‑task long‑term persistent memory; no built‑in knowledge base or RAG (Retrieval‑Augmented Generation) capabilities.

3.3 MCP Connector Ecosystem Baseline
• 34 built‑in official connectors, providing 151 standard tools:
◦ Local tools: file read/write, PDF/Word parsing, local Git retrieval, Shell commands, static web page scraping.
◦ Office & collaboration: Slack, Gmail, Outlook, Google Calendar, Notion, Confluence.
◦ Development tools: GitHub, GitLab, Jira, Linear.
◦ Business tools: Asana, ClickUp, HubSpot, Google Drive.
• Extension specification: any service implementing the MCP protocol can be plugged in seamlessly; developers may build custom internal enterprise connectors.
• Constraint: no connector marketplace is provided; adding new connectors requires manual configuration or custom development.

3.4 Automation and Templating Capabilities
• Cron‑expression scheduled recurring tasks – e.g., weekly reports, daily stand‑up briefings can be auto‑generated.
• Frequently used instructions can be saved as task templates for one‑click launch.
• Constraints:
◦ No file‑change or Webhook event‑driven triggers.
◦ Scheduled tasks terminate when the client application is closed.

3.5 Deliverable Output Baseline
•  Native output formats: Markdown, HTML, plain text, and structured tables, saved directly to local disk.
•  Outputs can be pushed to IM or email via connectors with a single click.
•  Baseline does not natively support direct generation of binary files such as Excel, PPT, or PDF – external conversion tools are required.

3.6 Security Baseline
• Directory allowlisting: restricts the local folders the Agent is permitted to access.
• Isolated credential storage for API/OAuth tokens; logs are sanitised to prevent plaintext exposure of secrets.
• Built‑in basic prompt injection protection.
• Gaps (baseline omissions):
◦ No native sandbox isolation.
◦ Single‑instance, single‑user – multi‑tenancy is not natively supported.
◦ Untrusted scripts pose local operational risks.

Ask for an outcome, not just an answer. OpenWorker works in the tools you use every day — your files, Slack, calendar, and more — and carries tasks through from start to finish: delivering a polished document, sending a Slack message, or updating a calendar entry. It checks in before important actions, so you stay in charge.

OpenWorker

IV. Capabilities Explicitly Excluded from the Baseline (Critical Boundaries for Technology Selection)

  1. Not designed for 24/7 unattended automation – cannot replace headless server‑side agents such as OpenClaw or Hermes.
  2. No distributed cluster scheduling – limited to single‑machine, single‑instance operation.
  3. No built‑in RAG or vector database – knowledge‑base capabilities rely on external MCP services.
  4. No standard external HTTP API gateway – cannot be directly used as an invocation endpoint for third‑party systems.
  5. No voice interaction support.
  6. No autonomous continuous learning, automatic tool discovery, or recursive sub‑agent creation.

V. Applicable Scenarios vs. Non‑Applicable Scenarios

Scenarios Aligned with Baseline Capabilities
• Desktop‑based human‑AI collaborative office automation for knowledge workers – meeting minutes, weekly reports, and information aggregation.
• R&D personnel integrating cross‑platform data from Jira/GitHub and automating document organisation.
• Enterprise customisation foundation – building private, local AI assistants for employees.
• Learning blueprint for Agent engineering – a complete reference implementation featuring TurnEngine, HITL, and MCP.

Scenarios Beyond Baseline Capabilities
• Long‑running cloud‑based IM bots for server‑side operations, or unattended O&M automation.
• Native Agent nodes for multi‑tenant, customer‑facing SaaS platforms.
• Large‑scale batch offline document retrieval and knowledge‑base systems.

VI. One‑Sentence Baseline Summary

Built on a Tauri+React frontend, a Python asynchronous backend, the TurnEngine agent runtime, the aisuite model abstraction layer, and the MCP tool protocol, OpenWorker v0.1.6 is a local‑first, model‑neutral, open‑source desktop AI Agent that runs on personal PCs, enforces a mandatory human‑in‑the‑loop safety mechanism, and delivers finished documents to knowledge workers—excelling at human‑AI collaborative office automation while remaining unsuitable for large‑scale, unattended server‑side automation tasks.

Official website: https://openworker.com
Open source code repository: https://github.com/andrewyng/openworker/releases

Leave a Comment