AI Bot Groupmate AI Agent

<div align="center"> <img src="https://gcore.jsdelivr.net/gh/Canvixel-Eloweny/ai-bot-groupmate@main/assets/logo-v5.svg" alt="Estrella Bot" width="140">

Estrella Bot · QQ-BOT-Creative

Turn a spare QQ account into an AI group member. It chimes in, picks up running jokes, lurks, and occasionally speaks up on its own.

OneBot v11 · NapCat · a local model or DeepSeek / Zhipu / Qwen · a zero-build web console

<p> <b>English</b> · <a href="./docs/README.zh-CN.md">简体中文</a> </p> <p> <a href="https://github.com/Canvixel-Eloweny/ai-bot-groupmate/actions/workflows/ci.yml"><img alt="Four-layer gates" src="https://img.shields.io/github/actions/workflow/status/Canvixel-Eloweny/ai-bot-groupmate/ci.yml?branch=main&label=four-layer%20gates&logo=githubactions&logoColor=white&style=for-the-badge"></a> <a href="https://github.com/Canvixel-Eloweny/ai-bot-groupmate/actions/workflows/codeql.yml"><img alt="CodeQL" src="https://img.shields.io/github/actions/workflow/status/Canvixel-Eloweny/ai-bot-groupmate/codeql.yml?branch=main&label=CodeQL&logo=github&logoColor=white&style=for-the-badge"></a> <a href="./LICENSE"><img alt="License" src="https://img.shields.io/github/license/Canvixel-Eloweny/ai-bot-groupmate?label=license&logo=gnu&logoColor=white&style=for-the-badge"></a> </p> <p> <a href="./package.json"><img alt="Platform" src="https://img.shields.io/badge/platform-macOS%20%7C%20Windows-2f77c4?style=for-the-badge"></a> <a href="./package.json"><img alt="Node" src="https://img.shields.io/badge/node-%E2%89%A522.13-339933?logo=nodedotjs&logoColor=white&style=for-the-badge"></a> <a href="./package.json"><img alt="Dependencies" src="https://img.shields.io/badge/dependencies-ws%20only-blue?style=for-the-badge"></a> <a href="./package.json"><img alt="Build" src="https://img.shields.io/badge/build-none-lightgrey?style=for-the-badge"></a> </p> <p> <a href="./README.md#4-interfaces-plugins-and-skills"><img alt="AI" src="https://img.shields.io/badge/AI-OpenAI--compatible-6f42c1?style=for-the-badge"></a> <a href="./docs/release/CHANGELOG.md"><img alt="Release" src="https://img.shields.io/badge/release-v0.1.9-e05d44?style=for-the-badge"></a> <a href="https://github.com/Canvixel-Eloweny/ai-bot-groupmate/commits/main"><img alt="Last commit" src="https://img.shields.io/github/last-commit/Canvixel-Eloweny/ai-bot-groupmate?logo=git&logoColor=white&style=for-the-badge"></a> </p> <hr> <p>_{<b>Console preview</b> · 控制台预览画面 — run overview, persona, glass lab and persona studio}</p> <p><img src="https://gcore.jsdelivr.net/gh/Canvixel-Eloweny/ai-bot-groupmate@main/assets/panel-preview-v3.jpg" alt="The console: run overview, persona, glass lab and persona studio" width="880"></p> <hr> <p>_{The English README is the default. The Chinese edition lives in <a href="./docs/README.zh-CN.md"><code>docs/README.zh-CN.md</code></a> and is authoritative when the two differ.}</p> </div> <p align="center">_{<a href="#1-what-this-is">1. What this is</a> · <a href="#2-why-estrella-bot">2. Why Estrella Bot</a> · <a href="#3-install">3. Install</a> · <a href="#4-interfaces-plugins-and-skills">4. Interfaces, plugins and skills</a> · <a href="#5-how-it-works">5. How it works</a> · <a href="#6-known-limitations">6. Known limitations</a> · <a href="#7-roadmap">7. Roadmap</a> · <a href="#8-contributing">8. Contributing</a> · <a href="#9-contact">9. Contact</a> · <a href="#10-license-and-credits">10. License and credits</a>}</p>

1. What this is

Estrella Bot is an OneBot v11 bridge that plugs a spare QQ account into any OpenAI-compatible model, plus a zero-build browser console. It is built to act like a group member, not a command bot: it follows the conversation, answers when addressed, sometimes joins in on its own, remembers people and running jokes, and goes quiet when that is the natural thing to do.

At a glance:

• Group-native chat — @-mentions, name aliases, probability-based interjections with cooldowns, sleep hours, message splitting and human-paced sending.
• Bring your own brain — a local model (MLX / QwenChat), DeepSeek, Zhipu GLM, Qwen, or any OpenAI-compatible endpoint.
• A real control room — start/stop the bot, switch models, edit the persona, manage memory and extensions, watch logs and usage, all from a local browser UI.
• Small and self-contained — a single-process bot, one runtime dependency (ws), no build step, no central service.

You supply three things: a QQ account, a channel (NapCat over OneBot v11, which is the default, or the official QQ bot API which needs no protocol side), and a model (a local endpoint or an API key). Source runs need Node.js ≥22.13; the Windows installer bundles its own runtime.

The two channels are mutually exclusive and the choice is made once at startup — switching needs a restart. NapCat is the default because it talks as soon as the protocol side is logged in; the official channel stays silent until you supply an AppID + AppSecret. The official channel is the compliant route: your account gets a "机器人" label, and in exchange no third-party client touches its login state. It also has hard platform limits (passive replies within 5 minutes / 5 per message in groups, no merge-forward, no message history lookup, identities are opaque openids), all listed in docs/OFFICIAL-CHANNEL.md. New and existing installs both start on NapCat; switch in the console under 运行 → 切换协议端 once you have credentials.

The name. Estrella Bot is the software name. The project began as qq-bot, was renamed twice, and keeps the GitHub slug ai-bot-groupmate on purpose so existing links and search results still land here. The full history is in CHANGELOG.md.

2. Why Estrella Bot

• It behaves like a member, not a reply machine. A dedicated decision layer decides whether a message deserves an answer at all — @-mention, alias, question shape, interjection chance with cooldown, sleep window, per-member speaking style. Replies are split and paced like human messages, and the bot never answers itself.
• Model freedom without lock-in. Four presets keep local and cloud brains separate; switching only materialises the chosen preset, so settings never bleed across providers. Any OpenAI-compatible endpoint can be added as a custom brain, with per-model capability switches and an automatic degradation chain when a provider fails.
• Local-first and private. The bot, console, memory and logs live on your machine. Local models never leave it; cloud calls go only to the provider you configure. There is no central account system or hosted service.
• Everything is operable by click. The console is a zero-build static frontend plus a local HTTP server. Start and stop the bot, scan the protocol-side QR code, edit persona and memory, inspect logs, usage and traces — no bundler, no rebuild after an edit.
• Safety is the default, not a setting. The allow-list fails closed (an empty list refuses to start), outbound text is built as message segments so CQ codes from the model cannot become instructions, self-replies are dropped, SSRF and browse paths are gated, and test sandboxes never touch a live bot.
• Extensible without becoming a framework. Plugins and skills share one host, one loader and one settings UI. Deterministic hooks and model-callable tools are two kinds of the same package, not two APIs.
• Cross-platform, with the trade-offs stated. Windows gets a native installer and needs no Docker; macOS uses a Docker container because NapCat's macOS QQ allowlist is stale. No hidden platform magic.
• Small footprint, strong gates. The runtime dependency is only ws, and the UI has no build step. Changes go through a four-layer gate plus mutation tests. Gate counts are not frozen in this README — npm run verify is the only authoritative source for the numbers.

3. Install

Every route ends at the same local console. Pick the one that matches your OS.

What you need

Piece	Notes
A spare QQ account	Use an alt account; third-party protocol implementations carry a risk of account restrictions
A channel	NapCatQQ over OneBot v11 (default), or the QQ official bot API (needs an AppID + AppSecret from the open-platform console). NapCat is not bundled — its license does not allow redistribution
A model	A local OpenAI-compatible endpoint, or an API key for DeepSeek / Zhipu / Qwen
Node.js ≥22.13	Source runs only. The Windows installer embeds its runtime
Docker Desktop	macOS / Linux container route only. The Windows installer needs no Docker

Windows — installer (recommended)

1. Download Estrella-Bot-Setup-<version>.exe from Releases.
2. Run it. On the protocol side page, either let the installer download NapCat, or point it at an already-unzipped NapCat.Shell folder. NapCat is not bundled.
3. Launch Estrella Bot from the desktop or Start menu. The console opens, and the protocol side starts in the background.
4. In the console's protocol page, scan the QR code with the spare account.
5. Fill in the allowed group IDs and the model API key, then click One-click start.

Notes:

• Quit your desktop QQ before scanning, and keep the protocol-side process alive afterwards. NapCat works by injecting into QQ NT, which is single-instance; this is the most common "it worked yesterday" cause.
• No Docker, no WSL2 and no separate Node install are required.
• Prefer a manual or portable setup? Follow USER-GUIDE §9.

macOS / Linux — Docker (recommended on macOS)

Docker exists on macOS to route around NapCat's stale QQ allowlist: the container bundles a matching Linux QQ, so your own QQ client is never touched.

git clone https://github.com/Canvixel-Eloweny/ai-bot-groupmate.git
cd ai-bot-groupmate
npm install
cp config.example.json config.json        # then set allow.groups
bash scripts/bootstrap.sh                 # start the container and print the WebUI login URL
node scripts/check-onebot.js              # verify the protocol side and list your groups

Start the console and click One-click start in the page; or run the console and the bot as two commands:

node panel/server.js    # console at http://127.0.0.1:8788
npm start               # bot only; the console can also start it

On macOS you can double-click Estrella-Bot.app instead: it starts the console, opens the page, then brings Docker and the container up in the background.

• Full container guide, including rollback: INSTALL-DOCKER.md.
• The native macOS route is blocked by NapCat's allowlist. Do not install QQ from a third-party mirror — the chat client holds every credential for that account.
• There is no macOS installer or DMG; macOS runs from the repository.

Manual / development route (any OS)

On Windows, install Node.js, QQ NT and native NapCat, open the WebSocket server on 3001 and the HTTP server on 3000, then:

npm install
copy config.example.json config.json

Start the console with scripts\Estrella-Bot.bat, or run node panel/server.js and npm start separately. Step-by-step instructions, including what to do when a step fails: USER-GUIDE · 用户使用说明.

First run

1. Sign in to NapCat with the spare account.
2. Put the group IDs you allow into allow.groups. An empty list with allowAllWhenEmpty:false makes the process refuse to start.
3. Set the model endpoint and key in the console, or in config.json (apiKeyEnv is safer than pasting a key into the file).
4. Open the console. Desktop builds write the URL to USER_ROOT/panel/.panel.json; source runs default to http://127.0.0.1:8788.
5. Click One-click start, then @ the bot or call it by name in an allowed group.

Verify the installation

If you run from source, npm test exercises the behaviour regression without needing QQ or a model, and npm run verify runs the full four-layer gate. The gate numbers in this README are not frozen; the output of npm run verify is the only authoritative count.

4. Interfaces, plugins and skills

There are four interfaces worth knowing — the protocol side, the model, the console API, and extensions.

Interface	Between	What it is
OneBot v11	QQ ↔ bot	WebSocket on 3001 and HTTP on 3000; NapCat in practice
Model	bot ↔ LLM	Any OpenAI-compatible /chat/completions endpoint
Console API	browser ↔ panel	Local HTTP on 127.0.0.1; every route is documented
Extensions	package ↔ host	Plugins and skills share one host and one loader

Console — the human interface

The console is served by panel/server.js on 127.0.0.1. It is the control room: run overview, start/stop, protocol login, persona studio, memory, scheduled messages, extensions, logs, usage, traces and settings. The frontend is static assets under panel/next/ with a data-driven control table (schema.js), so there is no build step. The server binds to an OS-assigned port in desktop builds and writes the actual URL to USER_ROOT/panel/.panel.json; development runs default to 127.0.0.1:8788.

More: docs/FRONTEND-V2.md · docs/MOTION-METHODOLOGY.md.

Local HTTP API — the automation interface

Everything the console does is a local HTTP route: state, config, bridge start/stop, protocol login, model switching, memory, extensions, logs and diagnostics. Inputs, outputs, side effects and idempotency for every route are in docs/API_CONTRACT.md. The frontend polls /api/state; external tools can do the same, or call specific routes. The panel binds loopback only and checks Host and Origin, so it is not exposed to your LAN by default.

Models — four presets, plus your own endpoint

The console organises model settings into four presets. Switching only materialises one preset into flat fields for the bridge; presets never bleed into each other, and per-brain edits survive switching away and back.

Preset	Name	baseUrl	Traits
local	Local model	http://127.0.0.1:<port>/v1 (MLX or QwenChat)	Offline, free, data never leaves the machine. The two local channels are mutually exclusive (same model memory)
deepseek	DeepSeek cloud	https://api.deepseek.com/v1	Cheap and good, text only
zhipu	Zhipu GLM	https://open.bigmodel.cn/api/paas/v4	The only preset with web search + vision; also the always-thinking family
qwen	Qwen (Alibaba)	https://dashscope.aliyuncs.com/compatible-mode/v1	Free grant per model (1M tokens, 90 days); the fallback chain is long on purpose. The grant runs out, and the platform's "stop when free quota is used up" switch is off by default — turn it on

Custom models — bring your own key. Any OpenAI-compatible endpoint works: a company gateway, a self-hosted relay, LM Studio on another machine, or a provider we have not wired up. Give it a name, paste the API address and key, fill in the model name, then tick what it can actually do:

Tick	What it changes
Vision	Pictures from the group get sent to it. Unticked, they become [图片].
Thinking	A reasoning parameter is sent. If the server rejects it, the request is automatically resent without it.
Web search	The search tool is attached. ⚠️ This is Zhipu's shape; most other endpoints will not recognise it.

Only the boxes you tick appear afterwards. You can keep as many custom models as you like, switch between them like any other brain, and right-click one to rename or delete it. The address must be the API address (the one serving /chat/completions), not the provider's console page.

Custom models are stored in llm.customBrains, one entry per brain, not in llm.presets. The single source of truth for model capabilities (thinking levels, web search, vision, degradation-chain eligibility) is src/model-caps.js; the entry point for vision is supportsVision().

Plugins and skills — the extension interface

plugins/ and skills/ share one host, one loader, one manifest contract, one tool registry and one hook bus. kind only decides the default semantics; it is not a second API and it is not a sandbox.

	Plugin (plugins/<id>/)	Skill (skills/<id>/)
Best for	Deterministic capabilities: hooks, guards, fixed-time actions	Model-facing capabilities: tools plus optional prompt guidance
Entry	setup / activate / hooks (at least one)	Same, or prompt.sections for a prompt-only skill
Execution	Guaranteed when a hook fires	The model decides whether to call the tool
Prompt	Optional extension guidance	Optional skillExt section
Enable list	custom.extensions.plugins.enabled	custom.extensions.skills.enabled
Console page	Factory → Plugin & Skill Factory (plugin mode)	Factory → Plugin & Skill Factory (Skill mode)

A minimal prompt-only skill is two files:

skills/group-etiquette/
  skill.json   # id / apiVersion / entry + prompt.sections[]
  index.js     # may be just: export {};

What both kinds share:

• One manifest contract (plugin.json / skill.json): id, apiVersion, entry, optional permissions.
• One tool registry and one tool cap; one hook bus with a per-hook timeout.
• One settings UI, with sensitive-key handling and range checks.
• One enable list per kind. A package sitting in the directory is scanned and reported, but not loaded until its id is in the allow-list.
• No hot plug: changing the allow-list requires restarting the bot.
• Failure is visible: load failures, missing hooks and unregistered tools are reported instead of silently ignored.
• Extension hooks are fail-open (a broken package should not silence the bot); the outbound credential gate is fail-closed.

Extensions are not sandboxed — they load in the bot process and can import fs. Install only packages you trust. The host limits them with a permission-trimmed API (fetch allow-list, read-only OneBot proxy), egress quota and manifest checks, but that is a boundary, not a jail. See docs/SKILLS.md and .github/SECURITY.md.

None ship by default. The packages written for this project were removed; drop your own into plugins/ or skills/, install a ZIP from the console, then enable it in the allow-list. Third-party packages you add are not committed (see .gitignore).

Do not confuse the on-disk skills/ packages with custom.skills in config.json: the former is executable code, the latter is prompt data.

5. How it works

The shape of the system

QQ alt ──► NapCat (OneBot v11) ──► src/index.js ──► model
             ws 3001 / http 3000     │ trigger → context → prompt → split
                                      ├─ local  : MLX / QwenChat
                                      ├─ cloud  : DeepSeek / Zhipu / Qwen
                                      └─ custom : any OpenAI-compatible endpoint
                 ┌────────────────────┴────────────────────┐
                 │ panel/server.js (127.0.0.1)              │
                 │  ├─ panel/next/ static UI, polls state   │
                 │  └─ start/stop, config, memory, extensions│
                 └──────────────────────────────────────────┘

Message lifecycle

1. src/onebot.js flattens a OneBot event into text, image URLs and "was I mentioned".
2. src/index.js applies allow/deny lists, sleep and wake rules, and rate gates.
3. src/brain.js decide() answers one question with a pure function: should this message get a reply at all?
4. src/context-budget.js assembles the prompt: stable sections first, volatile context last, within a character budget.
5. src/llm.js calls the model with the degradation chain and explicit 400 handling.
6. src/tool-loop.js runs tool rounds when the model asks for a tool.
7. src/brain.js splits the reply and strips internal markers; src/egress.js runs the outbound gate.
8. src/pace.js sends at a human pace and re-checks the allow-list before each send.

Configuration

cp config.example.json config.json   # config.json is pre-filled and gitignored

Minimal working config (local MLX):

{
  "llm": {
    "baseUrl": "http://127.0.0.1:8080/v1",
    "model": "qwen3-8b",
    "localChannel": "mlx",            // mlx / qwenchat
    "thinking": { "mode": "off", "level": "low" }
  },
  "allow": { "groups": [100000001], "private": [], "allowAllWhenEmpty": false },
  "persona": { "name": "小鱼", "file": "persona/qq-chat.md" }
}

Cloud DeepSeek — keep the key in the environment, not in the file:

{ "llm": { "baseUrl": "https://api.deepseek.com/v1", "apiKeyEnv": "QQBOT_API_KEY", "model": "deepseek-flash" } }

export QQBOT_API_KEY=sk-xxxx
node src/index.js

Key options:

Field	Meaning
allow.groups / allow.private	Allow-list. If empty while allowAllWhenEmpty:false, the process refuses to start
allow.allowAllWhenEmpty	true hands the account to the model — local debugging only
deny.groups / deny.users	Deny-list; takes precedence over the allow-list
trigger.requireAtInGroup	Whether the bot must be @-mentioned or called by name in groups
trigger.aliases	Names that count as being addressed (e.g. ["小鱼"])
trigger.interjectChance	Probability of chiming in when not addressed (0–1); default 0
trigger.interjectCooldownMs	Cooldown between unmentioned interjections
reply.splitToken	Token the model uses to separate messages; default ||
reply.sendDelayMs	Delay between split messages, to avoid rate limits
throttle.globalConcurrency	Set to 1 for local models — concurrency doubles latency on Apple Silicon
throttle.perMinutePerSession	Per-session replies per minute
llm.presets.<key>	Per-brain settings; materialised on switch, do not edit the semantics by hand
custom	Workspace data (persona, enhancements, skills, memory, scheduled messages, play rules, security, other)

Persona and memory

Persona is plain natural language in persona/qq-chat.md. buildMessages() in src/brain.js assembles the prompt in this order:

style rules → persona → enhancements → play rules → group background → context → current message

To make replies livelier, in order of value: write a concrete persona (catchphrases, message length, when not to reply) · add few-shot examples inside the persona file · lower temperature (≈0.8 is stable; above 1.2 it starts making things up) · use a bigger model.

The custom workspace is data-driven; its structure lives in src/custom-config.js. Partial updates go through patchCustom(), which merges instead of replacing, so editing one field cannot wipe the others. Persona changes invalidate stale session history via a fingerprint; memory is deliberately excluded, so adding a memory does not clear history.

Repository layout

Path	What lives there
src/	The bot: protocol IO, trigger decision, prompt assembly, model client, tools, memory, extensions
panel/	server.js (local HTTP API) + next/ (zero-build frontend) + lib/ (backend layers)
plugins/ · skills/	Extension packages — none ship by default
scripts/	Bootstrap, OneBot probe, dry run, event sniffing, contract gates
test/	Behaviour regression, sandbox, panel boot/DOM checks, mutation tests, E2E scripts
packaging/	Windows NSIS installer, portable build, update manifests and signing
docs/	User guides, API contract, architecture, skills, legal notices
persona/	Persona text
config.example.json · docker-compose.yml	Sanitised config template and NapCat container

Layering and dependency direction are explained in docs/ARCHITECTURE.md.

Design decisions

• CQ injection defence — everything sent to the protocol side is a message-segment array, never a CQ code string. If the model emits [CQ:at,qq=all], it stays ordinary text.
• No self-replies — events where user_id === self_id are dropped, otherwise the bot answers itself forever.
• Allow-list fails closed — an empty allow-list without an explicit opt-in refuses to start, rather than silently letting everything through.
• Reconnect with exponential backoff (1s→30s); in-flight requests are rejected on disconnect, never left hanging.
• Rate limiting — per-session minimum interval, per-minute cap, global concurrency semaphore.
• Long-message splitting prefers sentence boundaries so the protocol side does not truncate mid-sentence.
• Degradation chain is validated on both ends — cross-provider, duplicate and over-long entries, and always-thinking models are rejected with a stated reason.
• Log ring buffer — state.logs is capped, the backend keeps a dropped counter so total stays monotonic, and a lagging frontend cursor triggers a reset redraw.
• Sandbox isolation — test/sandbox.sh starts the panel with QQBOT_SANDBOX=1; in that mode the server never adopts or stops a real bridge process, so regression runs cannot kill your live bot.

6. Known limitations

• No streaming — the whole reply is generated before sending. Character-by-character output in a group chat looks less human, not more.
• Voice and video become placeholder text. Images are genuinely understood by vision-capable models; text-only models still receive a placeholder.
• Cross-restart context relies on session archives, and archives past a freshness threshold are refused — better to forget than to resume a topic that ended long ago.
• The bot is single-process and single-account. Multi-account is not implemented.
• The console reads from disk on every request (zero build). Editing server.js requires restarting the panel process (POST /api/panel/restart is idempotent).
• Extensions can execute, but "on disk" and "in effect" are two different things — packages are scanned and listed, but only ids in the allow-list are loaded, and there is no hot plug.

7. Roadmap

Everything below is known-not-done, taken from the limitations above. No time commitments — this is a one-person project and priorities follow real pain.

• Multi-account support — currently single process / single instance.
• Understanding voice and video content — today they become placeholder text.
• Outbound rich media — sending images. (Sticker output was removed: the bot only ever picked from QQ's built-in faces.)
• Native macOS protocol side — blocked by NapCat's QQ version allow-list.
• Hot-reloading extensions — changing the allow-list still requires restarting the bot.

8. Contributing

Start with .github/CONTRIBUTING.md. Three things to know before writing code:

1. The gates have four layers — check-wb structural contracts, smoke behaviour regression, and the sandboxed presets and panel self-checks. Run the whole thing with npm run verify.
2. Every fix needs a mutation test that only catches that one fix. "The assertion exists" is not the same as "the assertion is wired up".
3. Gate counts are deliberately not hard-coded in this document. They rotted four times before anyone noticed; the real numbers live in exactly one place: the output of npm run verify.

Also read docs/release/CHANGELOG.md, .github/SECURITY.md (including the known design boundaries), .github/CODE_OF_CONDUCT.md, and docs/legal/THIRD-PARTY-NOTICES.md.

9. Contact

Channel	Use
Issues	Preferred. Questions, bugs, feature requests — public discussion is searchable by the next person; a DM is not
.github/SECURITY.md	Report vulnerabilities privately; do not open a public issue

Two things to do first: redact real QQ numbers / group ids / chat logs before posting, and run npm run verify and paste the output — a minimal reproduction beats a long description.

10. License and credits

Item	Value
License	Apache-2.0 — full text in LICENSE, attribution notice in NOTICE
Third-party notices	THIRD-PARTY-NOTICES.md
Runtime dependencies	Only ws (MIT). Zero build, no other third-party runtime dependency

Apache-2.0 in three sentences. ① Use it freely: read, run, modify, ship it — even closed-source, no permission needed. ② Keep LICENSE and NOTICE with it and mark the files you changed. ③ It grants an explicit patent license and reserves trademark rights; it does not require you to publish your changes (there is no network clause).

NapCat is not part of this repository. Running it requires an OneBot v11 implementation (in practice NapCatQQ), whose custom Limited Redistribution License is not an OSI-approved open-source license and explicitly restricts building other projects from its code. This project therefore only calls it through official channels, shares no code with it, and distributes none of its code or binaries. Obtain it yourself and follow its license.

plugins/ and skills/ ship no extensions by default: the packages written for this project were removed. The host still loads any package you drop into either directory, subject to the allow-list in custom.extensions.plugins.enabled / custom.extensions.skills.enabled (the legacy custom.plugins.enabled is derived for old configs). There is no hot-plug: changing the allow-list needs a restart (the console says so). See docs/SKILLS.md for the two kinds. Also not distributed: napcat/ (the protocol side) and the data directories extensions write at runtime — see docs/legal/THIRD-PARTY-NOTICES.md §3.

---

<div align="center"> _{<b>English</b> · <a href="./docs/README.zh-CN.md">简体中文</a>} </div>