Two months ago I tried the MCP service of a domestic patent database vendor. Their AI offering seemed to have been out for a while, the docs looked complete and the feature coverage was broad, so I asked them for a trial. In practice the product is still rough, and it is not cheap either — in the end we simply had the agent drive a browser to reach the AI.

After I handed my workspace over to agents in May, my token usage went completely off the rails (see [My Token Usage Exploded 13× in Three Months]). Then DeepSeek raised its prices in August, the official API became genuinely unaffordable, and the only thing left to do was hunt down cheap third parties. So over the past six months I’ve tried pretty much every token plan (monthly or prepaid quota) I could get my hands on — three Chinese cloud vendors and five overseas ones.

This isn’t a recommendation list. It’s just a record of what I ran into: the good parts praised, the painful parts reported as they were.

A while back I said to a friend, half joking: “With the amount of AI I’m burning through, if I ever did get to be a boss, I’d squeeze my underlings harder than any boss I’ve ever had.” Then I got a little curious myself — is there actually any data behind that feeling?

So I pulled every commit and PR timestamp from 2025 and 2026 across all my repositories (40+) with the gh CLI, and looked at the distribution from two angles: volume and timing. Let’s see how black-hearted I really am.

I recently watched two videos — one about Vibe Coding and another about AI’s greatest help. Looking back, both explore the same question: what can Agents really change, and what do they bring?

DeepSeek recently officially raised its prices. As someone whose token usage was closing in on 100M tokens last quarter (see [My Token Usage Exploded 13× in Three Months]), I didn’t pay much attention at first — after all, DeepSeek was already very cheap, and even the raised price was still lower than the official price announced when V4 first launched. But I underestimated my usage: after a day of real-world testing at the new prices, it became clear that the cost is genuinely hard for an individual to bear. If I didn’t find a way around it, I’d be working just to pay for tokens.

When I handed my workspace over to Multica back in May, I was happily picturing how letting AI take on more of the work would mean fewer late nights for me. The work did get done — but my token bill went through the roof. The other day I pulled a three-month usage report for the workspace out of curiosity, and I had to double-check that I was reading it right: about 38M tokens in May, jumping to nearly 500M in July, with a three-month total of 883M. So here’s a data-driven look at how my token consumption exploded over those three months — and why. Because my AI tasks have gone from “trying it out” to “can’t live without it”.

After previously tinkering with putting Claude Science into a Docker container, I’ve been using it relatively intensively over the past two weeks (connected to DeepSeek, not the original Anthropic version, of course), so I want to talk about its advantages and limitations.

I recently worked on a small project: putting the Linux version of Claude Science into a Docker container, paired with an API Bridge so it can run through third-party API backends without needing to hold an Anthropic API Key directly.

The process was more eventful than expected—also quite interesting. Here’s a record of it.

Lately I’ve been working on integrating several AI applications, which means I need to switch between different models. That’s when I discovered a problem — the APIs of OpenAI and Anthropic are incompatible.

I had been thinking that when my blog’s total page views hit 50,000 again, I’d write a celebratory post. Little did I expect that the clicks would grow much faster than I imagined — before I could even finish drafting the celebration, the number had already blown past it.