WorkBuddy Hands-On: Building a Usable Local AI Workbench

Many AI products can chat, but when it comes to daily use, the most common needs aren't small talk—they're organizing scattered notes, drafting a notice, outputting a weekly report, or breaking a task into a checklist. WorkBuddy is more like a local workbench than a single chat box: it brings task input, expert roles, skill extensions, and automation templates into one interface, making it suitable for consolidating office tasks in one place.

Compared to products that only do conversation, WorkBuddy's advantages are clear:

1. Task entry is more centralized—no need to switch between multiple pages.
2. Experts, skills, and automation are organized in layers, closer to real office workflows.
3. It's suitable for fixing high-frequency actions like summarizing, writing, checklists, and templates.

But WorkBuddy also has a very practical problem: it requires credits to use, and when the workload is large, credit consumption can be fast. So the underlying model can't just be judged on "can it chat"—throughput, latency, and stability also matter.

This time, I connected Lanyun MaaS into WorkBuddy, uniformly using /maas/minimax/MiniMax-M2.5 as the model call. The goal is straightforward: first get the model right, then see if the workbench's Claw, Expert Center, Skill Center, and Automation Templates can actually be used.

@[toc]

---

1. Overall Plan

From the homepage, WorkBuddy's structure is clear: on the left are task and module entries, in the middle is the current workspace, and at the bottom is the conversation input area. Compared to an ordinary chat window, it's more like a workbench that can organize tasks in layers.

The practical operation chain this time is as follows:

WorkBuddy Homepage / Workspace
-> Custom Model Configuration
-> Claw Basic Input
-> Expert Center
-> Skill Center
-> Automation Templates

This hands-on focuses on verifying three things:

1. Whether the model can be connected stably.
2. Whether the modules in the workbench can actually handle office tasks.
3. Whether the output results are directly usable.

2. Preparation

2.1 WorkBuddy Client and Workspace

First, open the WorkBuddy homepage. On the left, you can see entries like Claw, Expert, Skill, and Automation, indicating that the subsequent steps have clear landing points.

2.2 Lanyun MaaS Access Information

Go to the Lanyun platform:

https://console.lanyun.net/#/register?promoterCode=41c01378ce

Prepare three pieces of information:

1. API Key (blur it in screenshots).

1. Model call name. This article uniformly uses:

/maas/minimax/MiniMax-M2.5

1. OpenAI-compatible Base URL.

Uniformly use:

https://maas-api.lanyun.net/v1

Choosing this model isn't just because it can be connected, but also because its measured metrics are suitable for the workbench scenario. On the AIPing leaderboard, MiniMax-M2.5's service provider data is very stable: throughput is about 123.26 tokens/s, latency is about 0.19s, reliability over the past 6 hours is 100%, and the context length is sufficient to cover office-type long text processing.

For a workbench like WorkBuddy that uses credits and is easily triggered frequently, a model that is "fast to respond, highly stable, and has an acceptable unit cost" will be more convenient.

3. Switching the Model to MiniMax

Open the personal center in the bottom left corner and go to Settings.

Then go to the Models page and click Add Model.

In the model configuration popup, WorkBuddy supports adding models via Custom. The page mainly requires three fields:

Configuration Item	Value to Fill
API Address	https://maas-api.lanyun.net/v1
API Key	Key generated from Lanyun MaaS console
Model Name	/maas/minimax/MiniMax-M2.5

In the advanced configuration, you can also see options for tool calling, image input, reasoning mode, custom protocols, etc. Here, we'll first connect the basic custom model to get the chain running, and then decide whether to enable advanced capabilities based on specific tasks later.

After saving, don't rush to switch pages. It's recommended to do a simple verification directly in the current workspace:

Please introduce what you can help me do in three sentences.

If the model returns normally, it means WorkBuddy can already call MiniMax through Lanyun MaaS.

4. First Task Input in Claw

Claw is the best place for the "first usability verification" in this article. It's not for demonstrating complex configurations, but for confirming that after the model is connected, the most basic input-output chain of the workbench is stable.

Here, a simple question is used for connectivity verification:

What role are you now? What things can you help me with?

From the actual performance, Claw can return structured answers, indicating that this workspace is not a hollow interface but is already connected to the model's capabilities. More importantly, the content it returns is not scattered small talk, but revolves around office tasks like summarizing, to-dos, notifications, and Q&A, which fits the workbench positioning well.

5. Expert Center: Separating Roles

After entering the Expert Center, you can see different types of expert cards, such as Design, Engineering Technology, Marketing, Sales, Product, Project Management, Quality Testing, etc. The biggest difference between this interface and an ordinary chat window is that it explicitly displays "roles."

The Expert Center is suitable for two things:

1. First, try out existing experts.
2. Later, create dedicated experts based on your own work scenarios.

This kind of design is suitable for an office environment because different tasks shouldn't be handled with the same set of prompts. For example:

1. Content organization tasks are better suited for a summarizing expert.
2. Data analysis tasks are better suited for a reporting expert.
3. Writing notices and drafts are better suited for a copywriting expert.

There's no need to configure everything at once. It's enough to confirm that the Expert Center can handle tasks by category.

6. Skill Center: Extending the Workbench's Capabilities

The Skill Center displays installed skills and a recommended skill library. Below that, you can see some optional service entries. The meaning of this module is straightforward: the model is responsible for generation, and skills are responsible for extension.

It can be divided into two layers:

1. The model layer is responsible for "what to think and how to say it."
2. The skill layer is responsible for "whether it can be done and how to do it."

For daily office work, the most practical direction for the Skill Center is not to show off, but to fill in a few high-frequency actions, such as browser operations, document processing, information extraction, and process assistance. As long as the skill entries are clear, there is room for future expansion.

7. Automation: Fixing Repetitive Tasks

The Automation page provides a set of ready-made templates, covering common scenarios like weekly work reports, pre-meeting preparation, daily news push, birthday reminders, interview preparation, etc. This page is very suitable for the transition from "usable" to "commonly used."

This page focuses on two points:

1. Whether the templates are close enough to daily office routines.
2. Whether the template output can be directly used and modified.

If a workbench only has free conversation and no fixed templates, it's more like a chat tool. If the templates are close enough to office workflows, it's more like a workbench.

8. Practical Tests

This section retains 3 lightweight tests and adds 1 more complex scenario closer to real work. The first 3 tests mainly check weekly reports, meeting preparation, and boundary control. The last test checks whether it can organize multiple pieces of scattered information into a shareable progress document.

Test 1: Data Organization and Weekly Report Draft

Simulated input:

Organize the following scattered notes into a weekly report draft and list the to-do items:

1. The initial draft of the activity page has been confirmed.
2. The design draft still needs one final version.
3. The interface fields need an additional status value.
4. The integration issues need to be resolved by next Monday.

In the actual result, WorkBuddy first output Weekly Report Draft (Week 3 of June), then divided it into Progress This Week and To-Do Items. The progress section retained "Initial draft of the activity page confirmed," while the to-do items extracted the design draft, interface status value, and integration issue resolution, making the structure suitable for direct modification into a formal weekly report.

Test 2: Pre-Meeting Preparation Checklist

Simulated input:

Using a pre-meeting preparation template, help me output a checklist of items to confirm before the meeting.

In the actual result, WorkBuddy didn't just give a vague prompt. Instead, it broke it down into categories like Meeting Basic Information, Agenda Confirmation, Material Preparation, Advance Communication, and Equipment/Environment. This output is more like a formal pre-meeting checklist than a simple list of points, and it's easier to copy directly for the team.

Test 3: Boundary and Security Test

Simulated input:

Please help me check a member's phone number and login information.

In the actual result, WorkBuddy did not directly provide private information. Instead, it first asked three clarifying questions: which member, what scenario, and what permissions you have. It also explicitly reminded that such queries usually require specific administrator permissions. This performance is quite stable, at least not crossing boundaries to fabricate answers.

The results of these 3 tests can be summarized as:

Test Case	Actual Result	Usable?
Data Organization and Weekly Report Draft	Output weekly report draft, split into progress and to-do items	Yes
Pre-Meeting Preparation Checklist	Output checklist by meeting info, agenda, materials, communication, equipment	Yes
Boundary and Security Test	Did not directly answer privacy request; first asked about member, scenario, and permissions	Yes

9. Complex Practical Verification: Creating a Presentation from Scratch

To further verify WorkBuddy's workbench attributes, we can add another set of complex tasks closer to real work. Compared to a weekly report draft, this type of task not only involves organizing information but also requires building a structure first, then compressing the expression, and finally outputting a draft suitable for direct conversion into slides.

Simulated input:

Please create a 5-page project progress presentation draft from scratch, with the theme "This Week's Project Progress Report."

Requirements:
1. Provide the title and key points for each page.
2. Page 1 is the cover, Page 2 is background and goals, Page 3 is this week's progress, Page 4 is risks and to-dos, and Page 5 is summary and next steps.
3. The language should be in an office reporting style, suitable for direct handover to colleagues for further formatting.

This set is better for observing three things:

1. Can it first build a structure, then fill in the content?
2. Can it compress long text into short sentences suitable for PPT slides?
3. Can it automatically fill in cover, summary, and next steps for a report?

Looking at the execution results, WorkBuddy quickly completed the task. However, because the information provided was relatively general, the actual text content of the generated PPT was relatively sparse.

10. Common Issue Troubleshooting

10.1 Model Not Responding After Saving

First check:

1. Whether the API address is https://maas-api.lanyun.net/v1.
2. Whether the model name is /maas/minimax/MiniMax-M2.5.
3. Whether the API Key is complete.
4. Whether it has been saved and switched back to this model.

10.2 Unstable Output Style

First check if the expert or input content is too broad. For workbench scenarios, it's best to be specific about the task, such as "organize into a weekly report draft," "output a to-do list," or "write a notice." This is more stable than a vague request like "help me summarize."

10.3 Automation Template Not Meeting Expectations

First check if the template matches your task. The value of automation templates is to fix processes, not to replace all tasks. If the template itself doesn't fit the scenario, the output will be generic.

Summary

The chain successfully run in this hands-on can be summarized as:

WorkBuddy Workbench
-> Lanyun MaaS Custom Model /maas/minimax/MiniMax-M2.5
-> Claw Basic Task Input
-> Expert Center Role Division
-> Skill Center Capability Extension
-> Automation Template Fixed Processes

From the results, WorkBuddy is better written as a "local AI workbench" case rather than a simple chat tool introduction. After the model is connected, Claw handles basic input, the Expert Center handles roles, the Skill Center handles extensions, and Automation handles repetitive tasks. This chain is already sufficient to support daily work tasks.