AskFriday

Vencord userplugin. Hover any Discord message → Ask Friday button → an LLM drafts a human-sounding reply (using the surrounding conversation as context) and drops it into your compose box for you to review, edit, and send.

Desktop / Vesktop only. Requests go through the plugin's native.ts (the Electron main process) to dodge browser CORS. Web Vencord has no native side and can't reach OpenAI/Gemini.

Features

• Message-hover button that drafts a reply to that message.
• Conversation-aware - feeds the recent messages around the target (before and after, when they exist) so replies follow the thread.
• Three providers: Anthropic (Claude), OpenAI (ChatGPT), Google (Gemini).
• Two auth modes:
  • API key (BYOK) - pay-per-token. Official and stable.
  • Local CLI - uses your subscription via the official CLI's own OAuth login (claude, codex, gemini). Safe and sanctioned (no token handling, no scraping). Slower (spawns a process) and counts against your subscription.
• Tone controls - human / casual / technical / friendly / professional / witty / concise / robotic / custom, plus length, emojis, match-language, and free-form extra instructions.
• Regenerate - click again to replace Friday's previous (untouched) draft instead of stacking a second one.

---

Install

Friday rides inside Vencord, which rides inside desktop Discord / Vesktop. You build Vencord from source with this plugin folder dropped in.

Prerequisites

• Node.js 18+ and pnpm (npm i -g pnpm)
• Git
• Desktop Discord or Vesktop (not the browser)

Quick build (script)

From this repo, run the helper - it clones/updates Vencord, copies the plugin in, and builds:

./build.sh                              # builds into ~/Vencord
VENCORD_DIR=~/code/Vencord ./build.sh   # or a custom location

Then skip to Step 2 to load the build. The manual steps below do the same thing by hand.

Step 1 - build Vencord with the plugin (both clients)

# get a Vencord source tree
git clone https://github.com/Vendicated/Vencord && cd Vencord
pnpm install --frozen-lockfile

# drop AskFriday into userplugins (COPY, don't symlink - a symlink to a path
# outside the tree breaks Vencord's @api/@utils path aliases at build time)
mkdir -p src/userplugins
cp -r /path/to/AskFriday/askFriday src/userplugins/askFriday

pnpm build

Step 2 - load it into your client

Pick the one you use.

Stock desktop Discord - inject:

pnpm inject     # arrow-key pick your desktop Discord, Enter

Fully quit Discord (tray included) and reopen.

Vesktop - do NOT inject. Point Vesktop at your build instead:

1. Open Vesktop Settings -> Vencord and find the Vencord location / custom build option.
2. Press Change and select your Vencord/dist folder.
3. Restart Vesktop.

Then in either client: Settings -> Vencord -> Plugins -> search "AskFriday" -> toggle on -> click the cog to configure.

Updating later: re-cp the folder, pnpm build, restart the client. No need to re-inject or re-point.

Vesktop + Flatpak note

If your Vesktop is the Flatpak build (~/.var/app/dev.vencord.Vesktop/ exists; flatpak list | grep -i vesktop lists it) it runs sandboxed:

• The Vencord/dist you select must be in a path the sandbox can read; grant it with flatpak override --user --filesystem=/path/to/Vencord dev.vencord.Vesktop.
• Local-CLI / subscription mode won't see your host claude/codex/gemini (they're outside the sandbox). Use API-key mode, or run a non-Flatpak Vesktop (AUR vesktop-bin, AppImage) for CLI mode.

The AUR/native build (~/.config/vesktop/; pacman -Qo $(command -v vesktop) names a package) has no sandbox - everything works as-is.

---

Configure

Open the AskFriday cog (Plugins page). Settings:

Provider & auth

Setting	What it does
Which AI to ask	Anthropic / OpenAI / Gemini
How to authenticate	API key or Local CLI (subscription)
API key	shown in API-key mode - stored plaintext, use a scoped key
Model	per-provider dropdown; Custom model id overrides it
CLI binary path	shown in Local-CLI mode - blank = claude/codex/gemini on PATH

Tone & shape

Setting	What it does
Reply tone	Human (natural) default - a helpful, knowledgeable engineer using correct technical terms. Also casual, technical, friendly, professional, witty, concise, robotic/bot-like, or custom
Custom tone	free-text tone (used when tone = Custom)
Reply length	short / medium / long
Allow emojis	on/off
Reply in the same language	mirror the original message's language
Extra instructions	appended to every prompt

Conversation context

Setting	What it does
Feed recent channel messages as context	on by default - gives Friday the thread around the message so replies follow the conversation
How many recent messages	slider: 0 / 5 / 10 / 15 / 20 (default 15) - the lead-up count; up to 5 follow-up messages are added automatically when the target isn't the latest

Context is read from Discord's in-memory message cache (whatever's loaded in the channel) - no extra network fetch. The model is told to use it only when the message is part of an ongoing conversation, otherwise reply on its own.

---

Using it

1. Hover a message in any channel.
2. Click the sparkle Ask Friday button in the hover toolbar.
3. Toast: "Friday is thinking…" → ~1–3s later the reply lands in your compose box (CLI mode is slower).
4. Not sent. Read / edit / regenerate (click again) → press Enter yourself.

Auth modes in detail

API key - paste a key for the active provider. Plaintext in settings; use a scoped/limited key.

Local CLI (subscription) - log the relevant CLI in first:

claude /login      # Claude Pro/Max
codex login        # ChatGPT subscription
gemini             # Google login

The plugin shells out to the already-logged-in CLI; it never reads a token. This is the safe way to use a subscription - not web session scraping, which violates ToS and risks bans (intentionally not implemented here).

---

Develop / test

Pure logic (prompt building, settings → request mapping) is Vencord-free and unit tested:

npx tsx --test askFriday/pure.test.ts

End-to-end provider smoke test with your own keys (runs the real native.ts calls, outside Discord):

# API-key mode
ANTHROPIC_API_KEY=sk-ant-... OPENAI_API_KEY=sk-... GEMINI_API_KEY=AIza... \
  npx tsx askFriday/smoke.ts

# subscription / local-CLI mode (log the CLIs in first)
FRIDAY_CLI=1 npx tsx askFriday/smoke.ts

Layout

File	Role
index.tsx	plugin def, hover button, context gathering, draft insert/replace
settings.tsx	definePluginSettings: provider, keys, models, tone, context
config.ts	settings store → GenerateOptions (pure, tested)
prompt.ts	buildPrompt() - persona/voice + conversation context (pure, tested)
native.ts	Node-side HTTP calls + CLI spawning (CORS-free)
providers.ts	shared types + model/CLI tables
pure.test.ts	tests for the Vencord-free modules
smoke.ts	live provider smoke test (uses your keys)

Notes / limits

• Context only includes messages already loaded in the channel's cache; Friday doesn't fetch history over the network.
• Provider model ids drift; the model dropdowns plus a custom-model override keep you unstuck.
• Local-CLI mode needs the CLI installed and logged in; errors surface as toasts.
• Personal userplugin - official Vencord doesn't accept AI/API-key plugins, so there's no upstream PR.