Office Raccoon 2.0: A Comprehensive AI Agent that Executes, Remembers, and Unifies Cloud and Local Workflows

Evolution line: execution → memory → usability

OpenClaw enables AI to run local commands, control browsers, and read/write files; its breakthrough is execution but requires manual configuration and has a high entry barrier.

Hermes adds memory and self‑evolution on top of execution, becoming more personalized with use, yet still needs curl installation and a self‑hosted server.

Office Raccoon Desktop 2.0 (by SenseTime) combines execution and memory, removes the shared setup friction, is polished for office use, integrates local and cloud capabilities, and provides a controllable mechanism.

Unified workflow in a single client

The client bundles two core abilities:

Local Claw operates inside the real computer environment—directly reads local files (20+ formats such as Excel, PDF, Word, CSV, PPT, images) with three‑level directory permissions, triggers the browser with a single sentence, and can be invoked globally via ⌘K / Ctrl K to connect Feishu or run scheduled tasks. This component handles execution .

Cloud Workbench leverages strong models for complex creation—data analysis, PPT generation, task planning, one‑image‑understanding, knowledge‑base Q&A, and copywriting. This component handles creation .

Consequently, a chain from “plan → analyze data → write report → generate PPT → execute” runs entirely within the same client, eliminating tool‑switching and data transfer between cloud and local environments.

High‑frequency capabilities of Desktop 2.0

Direct local file access without uploading. Granting a folder of monthly reports allows a single prompt such as “read, summarize key data by month, flag anomalies” to read, organize, and return results across Excel, CSV, PDF, Word, PPT, and images, respecting workspace/home/custom directory scopes.

One‑sentence browser control. Built‑in automation manipulates the local browser without plugins or page copying, enabling tasks that normally require RPA tools, especially for systems lacking APIs.

Quick Bar global invocation with write‑back to Excel. Pressing ⌘K (macOS) or Ctrl K (Windows) pops up a panel that can translate, summarize, or rewrite selected content, and can write analysis results directly beside the data in Excel.

Scheduled tasks. Users can configure daily pushes (e.g., news briefs) that run automatically each morning.

Integration and memory across data sources

The assistant integrates with Feishu (with enterprise WeChat and DingTalk support forthcoming), allowing AI results to be saved as Feishu docs, appended to existing docs, or searched in the knowledge base, moving output from a chat window into team collaboration flows.

Memory spans three data sources—cloud knowledge bases, local files, and third‑party services (e.g., Feishu)—so the assistant’s “memory” lives across cloud, edge, and external services without requiring a self‑hosted server.

Real‑world test scenarios

One‑sentence report generation : Providing a dataset lets the assistant analyze brand and model trends, produce charts and conclusions, and compile a “New Energy Vehicle Sales Analysis” report.

One‑click PPT conversion : The generated report can be turned into a PPT, with each slide editable via dialogue and exportable as a .pptx file.

Comprehensive feature set

Pre‑defined roles (industry researcher, data analyst, PPT designer, Feishu document assistant, chief writer) can be selected without manual prompt engineering. Skills cover common office formats (pptx, docx, xlsx, pdf) and are extensible.

Under the hood is a dual‑engine architecture with offline local models (Ollama / llama.cpp), an MCP tool ecosystem, and one‑click AI file rollback, providing a full stack of modern agents packaged in a hassle‑free client.