DeepSeek Bot: 1,400 Stars in a Week

Verdict: DeepSeek's official consumer agent ("Muse") is still nowhere in sight. The community didn't wait: DeepSeek Bot, an open-source plugin that turns DeepSeek Harness into a team of persistent bots with long-term memory, pulled 1,409 GitHub stars in under a week — roughly 200 people a day pressing the button. It's Apache 2.0, it runs on models you already have in DSH, and it costs nothing but tokens. Just don't expect production polish from a v0.1 that's six days old.

For context: I wrote a DeepSeek Harness quickstart back in September, when DSH was mostly interesting as a coding-agent harness. A bot-team plugin changes what DSH is for. This is worth unpacking.

What DeepSeek Bot actually is

The plugin (repo: Fei2411/DeepSeek-Bot, currently at v0.1.3) adds "bots" to DeepSeek Harness. Each bot is a DSH session with its own identity — name, role, instructions, and whatever model you point it at, since DS Bot reuses the models already configured in DSH. Bots persist: their state, memory, and chat history survive restarts, stored as Markdown files in your local DSH folder under bot/ and sessions/.

Three building blocks:

• Bots — persistent agents with long-term memory. Each round of conversation triggers a memory review pass: the model itself decides what to save, update, or drop. Oversized conversations get compressed and handed off via summaries.
• A main bot ("Chief") — a high-privilege administrator. You talk to Chief; Chief configures the other bots and sets up group chats. The author strongly recommends configuring everything through conversation rather than clicking through settings.
• Group chats — multiple bots in one thread, with bot-to-bot messages.

Everything is a plugin, true to DSH's design philosophy. The npm packaging comes from the official @deepseek-ai/dsh namespace, but DS Bot itself is a third-party community project — one JavaScript developer, not DeepSeek.

The token problem, and the admin fix

Here's the part I find most honest about this project: multi-agent group chats are token furnaces. Every bot that reads the thread burns input tokens; every reply burns output. Anyone who has run a "team of agents" on an API meter knows how fast this compounds.

DS Bot's answer: in a group chat, an admin bot answers by default and calls in other bots only when needed. You can @ a specific bot to bypass the admin. That's not a glamorous feature, but it's the right engineering call — it caps the fan-out that makes group chats expensive. If you're running bots on DeepSeek's API, pair this with context caching, since persistent bots replay a lot of stable context every turn.

How to install it

DSH Desktop (Windows/macOS): sidebar → Plugins → Add plugin → enter ds-bot as the package name → Install → Enable. Mainland China users should switch the registry to the China mirror. If DSH says changes apply after restart, fully quit and reopen.

DSH Web (Windows/macOS/Linux): needs Node.js 22.19 or later plus pnpm:

npx @deepseek-ai/dsh@0.2.1-alpha.1 plugin --profile web add ds-bot
npx @deepseek-ai/dsh@0.2.1-alpha.1 web

If npm is unreachable, the Releases page offers a ds-bot-0.1.3.tgz you can install by full path.

My take: what this says about DeepSeek's agent play

Two things, and I'll say them plainly.

First, the speed. The repo was created on October 7. Chinese tech media covered it around October 11 when it had roughly 250 stars; as I write this on October 13 it's past 1,400, with the last commit pushed this morning. Grok's bot feature is the obvious comparison point in the conversation — "Grok Bot costs too much to keep alive, so let's try the DeepSeek one" was literally the angle of the Chinese coverage — and an open-source clone that runs on whatever model you already pay for is a genuinely different cost proposition.

Second, and more interesting to me: DeepSeek's agent ecosystem now has two layers moving at different speeds. The official layer (Agent Integrations, Harness itself) advances deliberately. The community layer just demonstrated it can ship a consumer-flavored product before DeepSeek announces its own. My prediction: before an official "Muse" lands, we'll see more of these plugins, and the ones that survive will be the ones that treat token economics as a first-class design constraint, like the admin-bot mechanism does here.

Now the honest limits. This is a v0.1 preview from a single maintainer, six days old. I haven't run a bot team for weeks, so I can't tell you whether the automatic memory review is genuinely reliable or just plausible — that's the kind of claim only time settles. The bot's own author recommends sticking to one model per bot and switching rarely, which tells you the multi-model path is not battle-tested. If you try it, treat your first week as an experiment, not a deployment.

Sources: Fei2411/DeepSeek-Bot (GitHub API + README, 2026-10-13); QbitAI, 2026-10-11 (250+ stars at first-day coverage, admin-bot mechanism description). No affiliate links; no compensation involved.