deepseek-bot

An always-on AI teammate for Slack, Discord and Telegram, built on DeepSeek Harness.
It remembers your threads, reads what you send it, drafts documents,
and asks before it deletes anything.

Sponsors These are our sponsors. They help keep deepseek-bot going as a personal side project.

	OpenGateway sponsors deepseek-bot. An OpenAI-compatible gateway that puts many model providers behind one API. Thank you for backing open source.

curl -fsSL https://raw.githubusercontent.com/rlaope/deepseek-bot/master/install.sh | bash
dsh                    # talk to it in the terminal
dsh gateway setup      # connect Slack, Telegram or Discord: ↑↓ and Enter

The first line installs it, and running it again updates. It needs macOS with Node.js 22+, the DeepSeek Harness app and a DeepSeek API key — none asked for if the harness already has one. When the app is missing, the installer offers to download it from DeepSeek's own server (a few hundred MB, into /Applications), checked against the build's published hash. DeepSeek's own API is the default and needs nothing else; OpenGateway or another OpenAI-compatible endpoint is an option the installer offers. The bot starts with the terminal, dsh.

dsh gateway setup connects a messenger: pick one with ↑↓ and Enter, and it opens the page where each token is made, checks each token as you paste it, and offers to restart the bot. Without a terminal it takes the tokens from the environment instead, so an agent can do it for you (docs/setup.md); dsh gateway status says whether it is up.

dsh update

Without a terminal the installer asks nothing: the language comes from LANG, the model route is DeepSeek's own, a key is used where the harness already has it (its app, ~/.dsh/.env), and the DeepSeek Harness app is downloaded only when told to:

curl -fsSL https://raw.githubusercontent.com/rlaope/deepseek-bot/master/install.sh | DEEPBOT_INSTALL_HARNESS=1 bash

Where to put the key first, and what each stop means: installing without a terminal.

Optional: faster on OpenGateway. It serves deepseek/deepseek-v4.1-flash-ultrafast on an OpenAI-compatible endpoint (https://apis.opengateway.ai/v1), which it lists at 12× the speed and 1/30 the price of a Fable-class model. This bot is developed against it; the model route in profile/cordis.patch.yml shows where the endpoint and the model go.

It lives where your team already talks. Mention it in a Slack channel, a Discord server or a Telegram group, or message it directly. Nobody has to open another tool to ask it something.

It has a terminal too. dsh opens the same agent fullscreen in a terminal, beside Slack or on its own: answers stream in, tool calls show their diffs and results, and approvals and questions are answered in place. See the terminal.

You can see it working. The moment it takes a message the platform's own indicator shows: Slack's working status under the thread, "typing…" on Discord and Telegram. The answer is the only notification you get.

It remembers. Every thread is a conversation that survives restarts. Ask what was decided last week and it finds the passage and says what it implies, drawing only on the channel you are in, so a private conversation never surfaces in a public one.

It has a character you write. Who it is, how it speaks and what it knows about you and your work are plain Markdown files in its home. Edit them and it changes; it keeps its own notes there as it learns.

It does the work, not just the talking. Ask for a report and it writes the docx, xlsx or pptx and uploads it into the conversation. Send it a link and it reads the page; send an image and, with a model that accepts images, it looks at it.

It asks before anything destructive. A command that deletes or overwrites stops and waits for the person who asked: press the button, or type allow or deny.

It works in a room of its own. Its commands run in a sandbox that can write only its own workspace, and cannot read the deployment's tokens, other agents' files or the stored conversations, nor reach your SSH agent or password manager. Text in a thread or an attachment that tries to give it orders is flagged before the model sees it.

It speaks first only when asked to. "Tell me when this changes" becomes a watch that checks on a schedule; "remind me on Friday" becomes a reminder.

It stays up. It runs as a supervised service: it reconnects on its own, is restarted when a probe finds it unhealthy, can alert you when it does not recover, and tells a conversation when a restart cut its turn short.

It is small enough to read. TypeScript with zero runtime dependencies, covered by unit suites and by scenarios run against a real model.

• Mention it in a channel, or send it a direct message. It answers in the thread.
• A follow-up sent while it is working waits its turn.
• stop ends the turn that is running.
• usage shows the thread's token totals.
• allow or deny, or the buttons, answers a request for approval.

• macOS only: the services are launchd jobs and the sandbox is Seatbelt. On Windows it builds, the terminal's local transport listens on a named pipe, dsh forwards to the harness CLI named by DSH_BIN or else to npm's, found beside its dsh.cmd on PATH, and the suites in CI's Windows job pass; the installer, services, sandbox and watches are not ported there. The gateway can be run by hand with the terminal only, and refuses every command for want of a sandbox (Windows, by hand). With no sandbox the tool gate is no boundary on Windows: it checks file tools, and only the sandbox holds commands off. Its .env check follows Windows' own walk there (.. cancelled before links, drive-relative paths, NTFS streams, 8.3 short names), checked against real writes on NTFS. Another user can open the pipe read-only: that gets nothing, and gives way to a terminal when the places are full.
• Recall matches words, not meaning. A paraphrase that shares no distinctive word can miss.
• A page that draws itself with JavaScript is read with a headless browser, one page per run; a site that needs a sign-in, through a session you keep for it (--login).
• Slack is used live. Discord and Telegram are tested against mocked APIs, not yet a live server.
• On Telegram and Discord anyone who finds the bot can message it. List who may: dsh gateway setup telegram --users <your id> (who may talk to it).
• Its built-in notices and button labels are in Korean; what it says otherwise follows its persona.
• DeepSeek Harness is a release candidate, and this project leans on its internals. Pin your version.

• Setup and operations: connecting a messenger (dsh gateway), creating the bot on each platform, running the services
• How it works: the design, and the evidence behind each claim
• Limitations · Changelog · Plan
• For contributors and agents: working rules · vocabulary

MIT. deepseek-bot is an independent project, not affiliated with or endorsed by DeepSeek. The whale and the wordmark are the DeepSeek logo's; the DeepSeek name and logo belong to their owner.