# DeepSeek Bot turns DSH sessions into a team with a chief and a group chat

> DS Bot is a plugin for DeepSeek Harness that turns sessions into named bots with memory, a main bot called Chief, and group chats where one bot hands work to another. Each bot is a DSH session drawing on your own model quota, and a local MCP command added by a bot still needs your approval.

**Fei2411/DeepSeek-Bot** — A team of long-lived bots for DeepSeek Harness, with main bots, bot-to-bot messages and group chats.

- Repository: https://github.com/Fei2411/DeepSeek-Bot
- Stars: 2,865 · Forks: 276
- Language: JavaScript
- License: Apache-2.0
- Published: 2026-10-16 · Updated: 2026-10-16 · Language: en
- Canonical page: https://hysenlabs.com/projects/fei2411-deepseek-bot

## Every bot is a DSH session, which is what it borrows

The acknowledgement section states the architecture plainly: "Every bot is a DSH session and uses the DSH tools and models." Nothing is reimplemented here. The value added is identity, memory and routing on top of sessions you would otherwise open by hand.

That also explains the install requirement. The README says to install DSH and set up at least one model first, because "DS Bot uses the models from DSH". There is no separate model configuration, no key management, and no endpoint to set. If DSH has no model configured, DS Bot has nothing to run on.

The plugin is versioned separately and aggressively. The badge reads version v0.1 preview, and the three most recent releases are v0.1.4 on 2026-10-13, v0.1.5 later the same day, and v0.1.6 on 2026-10-14. Two releases in a day suggests a young project iterating on its own surface, so expect the plugin and your DSH version to move together.

The DSH side is pinned to `0.2.0-rc.2`, a release candidate. That is the harness, not the plugin, and it is the single most likely source of friction when you install.

## Two install paths, and the desktop one may need a restart

On DSH Desktop for Windows and macOS, installation happens in the sidebar: select Plugins, then Add plugin. In Package name or address you enter `ds-bot`, and in mainland China you set Registry to Mainland China mirror. Then you press Install and Enable now.

The next sentence is the one that catches people: "If DSH says the change takes effect at the next start, quit DSH completely and open it again." A window close is not necessarily a quit on macOS, so if the plugin does not appear, that is why.

If npm is unreachable, step 2 has two alternatives: the full path of `ds-bot-0.1.6.tgz` downloaded from Releases, or `github:Fei2411/DeepSeek-Bot#v0.1.6&path:/ds-bot`, which needs Git installed. Note that the tarball alternative pins a filename, so it goes stale as soon as a newer version ships.

The web variant covers Linux as well and has different prerequisites, Node.js `22.19` or later and pnpm:

```sh
npx @deepseek-ai/dsh@0.2.0-rc.2 plugin --profile web add ds-bot
npx @deepseek-ai/dsh@0.2.0-rc.2 web
```

Both lines pin the harness version explicitly, which is what you want when a release candidate is in play.

## Chief is a bot with rights, not a settings page

The first time you open DS Bot there is already a main bot called Chief, and talking to it walks you through the setup. The README's summary of its role is blunt: "In short: have a question? Ask the main bot."

The reason it can do that is a privilege difference, stated as: "The main bot is your chief. It has very high rights: it manages the other bots and the group chats." So configuration of the team itself is delegated to a bot rather than to a settings form. The README recommends exactly this and recommends strongly: "I strongly recommend that you set up your other bots and group chats by just talking to a bot."

That recommendation is also the main thing to think about. An interface that lets a bot create and manage other bots is faster to start and harder to audit. Nothing in the README describes a confirmation step for Chief creating a new bot, and the permission gate it does document is elsewhere: a local MCP command a bot adds needs your approval.

Individual bots also have a details page showing their info, where you pick the model each one uses. You can switch models at any time, but the README recommends keeping one model per bot for a long time and switching only when you really need to, which is advice about output consistency more than about cost.

## A group chat has an admin, and it costs more than one window

Group chats are where the multi-agent part becomes concrete. The README is direct about the price: a group chat "is a great way to get several bots to work together, but it clearly costs more".

The structure is that each group has an admin. The admin answers your messages by default and calls in other bots when needed, so you do not have to address each bot yourself. You can also @ a bot directly by typing @ and picking from a list, and it will answer you directly. To give bots another chat as context, the message box has a + → Reference a chat control.

Handoff is the mechanism that makes it a team rather than a chat room. Any bot can hand work to another bot with @, the receiving bot works on its own while "keeping the group posted", and the group, the sidebar and its own chat all show what it is doing. When it finishes it posts the result, and the bot that asked for the work answers you. To stop it, you open the group's details.

That last detail is the operational one. Cancellation is not a spoken command to the working bot; it is an action in the group's details page, so interrupting a runaway task means going to the UI.

## Bot memory is Markdown and you can edit it by talking

Bots have memory: a bot "remembers what you like and uses it in later work". The README says you can manage that memory on the bot's details page, or simply tell the bot.

The format is credited in the acknowledgements to AgentMemoryRepo, which published the Markdown format of bot memory. The plugin does not contain that project's code, and the acknowledgements list it among the projects whose designs informed DS Bot rather than among dependencies. Other listed influences are OpenAI Codex for checkpoints that keep user messages next to the summary and for message anchors used by chat search, OpenMausBot for group chat rounds and @mention relays, Hermes Agent and Hermes Bot Mode for one session per member per group chat and handoffs carrying earlier user messages, and OpenCode, goose, pi and Gemini CLI for when to condense a long chat and what a handoff note holds.

The Codex detail is the useful one for a user. If checkpoints keep your original messages next to the summary, you can read back what a bot condensed on your behalf, which is the difference between a memory you can audit and one you cannot.

Editing memory by conversation is convenient and imprecise. If a bot has remembered something wrong, the details page is where you fix it deterministically, and the README points at both routes without ranking them.

## Capabilities are shared, and only you or a main bot can change them

Settings → Capabilities is described as a shared library of MCP servers, skills and packs. Adding a capability to a bot in its settings makes it available from that bot's next message onward, in its own chat and in group chats.

The permission model is the part worth understanding. Three rules are stated. A local MCP command that a bot adds needs your approval before it can run. While other bots are using a capability, only you or a main bot can change or delete it. And the library is shared rather than per-bot, so a capability added for one bot is in the pool for the others.

The combination is a reasonable middle ground: bots can extend themselves, but not without you, and nobody can pull a capability out from under a running task. The gap is that approval is per command, not per capability, so a bot that adds a local MCP server asks you about each command it later wants to run.

Usage accounting sits in Settings → Usage, broken down by time, by bot and by model. Since every bot draws on the same DSH model configuration, that page is the only place to see which bot spent what, and with which model.

## Data stays in ~/.dsh and survives an uninstall

The Data section is one paragraph and has an important edge in it. "All data stays on your own computer, in `bot/` and `sessions/` in the DSH home folder (`~/.dsh` by default)." Then: "Uninstalling the plugin does not delete this data."

That is a deliberate choice with two consequences. Uninstalling does not clean up, so removing the plugin leaves bot definitions, memory and sessions on disk, and if the reason you are removing it is that you no longer trust it, the cleanup is a manual step. Conversely, an upgrade path that leaves state alone is one less thing to break.

The disclaimer section states the other half of the bargain: "Bots use your own model quota, and they read and write files and run commands when you ask them to. Check important actions yourself." File access, command execution and quota spend are the three things that make a multi-bot setup cost more than one window, and all three are attributed to the models you already configured.

The project also states plainly that it is an independent community project, not an official DeepSeek product, and that DeepSeek and DeepSeek Harness are trademarks of their owners used only to say what the plugin works with. The licence is Apache 2.0, provided "as is", without any warranty, with third-party code listed in NOTICE.

## Conclusion

Adopt DS Bot when you already run DeepSeek Harness with models configured and you keep bouncing between several sessions for one project, because a named bot with memory is a better handle for that than a session list. Do not adopt it expecting managed permissions or a team that verifies its own work, since the group chat is an orchestrating bot calling other bots and the README asks you to check important actions yourself. Before installing, confirm the DSH version you run matches the `0.2.0-rc.2` the instructions are written against, because the plugin is pinned to a release candidate and the desktop install path also needs DSH fully restarted for a plugin change to take effect.

## FAQ

### How do I install the DeepSeek Bot plugin?

On DSH Desktop, open the sidebar, select Plugins then Add plugin, enter `ds-bot` as the package name, then Install and Enable now, quitting DSH completely if it says the change applies at the next start. If npm is unreachable you can give the path of a downloaded `ds-bot-0.1.6.tgz` instead.

### Do bots need their own model configuration or API key?

No. You install DSH and set up at least one model there, and DS Bot uses the models from DSH. Each bot picks its model on its details page, and you can switch models at any time.

### How do I stop a bot that is working on a task?

To stop the work, open the group's details. Handoff through an @ mention is not interrupted by talking to the working bot.

### Where is bot data stored and does uninstalling remove it?

Data stays on your own computer in `bot/` and `sessions/` inside the DSH home folder, `~/.dsh` by default. The README states that uninstalling the plugin does not delete this data.

### Can a bot run a local MCP command without my approval?

No. A local MCP command that a bot adds needs your approval before it can run, and while other bots are using a capability only you or a main bot can change or delete it.

### Is DeepSeek Bot an official DeepSeek product?

The README states it is an independent community project, not an official DeepSeek product, and not affiliated with, endorsed by or sponsored by DeepSeek. DeepSeek and DeepSeek Harness are trademarks of their owners.

## Sources

- [Fei2411/DeepSeek-Bot on GitHub](https://github.com/Fei2411/DeepSeek-Bot)
- [Issues](https://github.com/Fei2411/DeepSeek-Bot/issues)
- [License: Apache-2.0](https://github.com/Fei2411/DeepSeek-Bot/blob/main/LICENSE)
- [README](https://github.com/Fei2411/DeepSeek-Bot/blob/main/README.md)
- [Releases](https://github.com/Fei2411/DeepSeek-Bot/releases)

---

Hysen Labs editorial analysis, written from the project's own repository and release notes. Cite the canonical page: https://hysenlabs.com/projects/fei2411-deepseek-bot