An autonomous X (Twitter) bot with a native desktop dashboard. The bot generates opinionated posts and replies based on RSS/web news feeds and your home timeline, powered by any OpenAI-compatible AI provider — local or cloud.
virtualMe/
├── service/ # Headless Bun process — runs 24/7, exposes REST + WebSocket
└── dashboard/ # Tauri v2 desktop app — connects to the service
- Autonomous posting — generates original posts from RSS feeds and web sources on a configurable schedule
- Autonomous replies — picks tweets from your cached home timeline and replies to them
- Thread support — automatically splits long content into reply chains
- Source URL injection — appends the news article link to posts (X t.co-aware, no truncation)
- DuckDuckGo context — enriches posts with background info fetched before the AI call
- AI prompt storage — every generated post stores the exact system + user prompt used, viewable in the dashboard
- Budget enforcement — hard limits on monthly X API calls (free tier: ~500 posts, ~100 reads); 5% reserve kept for manual triggers
- Fully configurable at runtime — all settings editable from the dashboard without restarting the service
- Live dashboard — real-time logs, post history with expandable prompts, news source management, budget tracking
| Dependency | Version | Notes |
|---|---|---|
| Bun | ≥ 1.1 | Both service and dashboard build tooling |
| Rust + cargo | stable | Required by Tauri for the dashboard |
| X Developer Account | — | Free tier is sufficient |
| AI provider | — | LM Studio, Ollama, or any OpenAI-compatible API |
- Go to developer.twitter.com and create a project + app
- Under User authentication settings, enable OAuth 1.0a with Read and Write permissions
- Generate Access Token & Secret (make sure they have read+write scope)
- Copy your numeric User ID from the app dashboard
Free tier limits: ~500 posts/month, ~100 reads/month. The default schedule (post every 2.5h, read timeline every 8h) stays well within these limits.
The service uses any OpenAI-compatible API. Set AI_BASE_URL, AI_MODEL, and AI_API_KEY accordingly:
| Provider | AI_BASE_URL |
AI_API_KEY |
|---|---|---|
| LM Studio (local) | http://localhost:1234/v1 |
not-needed |
| Ollama (local) | http://localhost:11434/v1 |
not-needed |
| OpenRouter (cloud) | https://openrouter.ai/api/v1 |
sk-or-... |
| OpenAI | https://api.openai.com/v1 |
sk-... |
For local models, any instruction-following model works. Reasoning models (DeepSeek-R1, etc.) are supported — the service handles the reasoning_content fallback automatically.
git clone https://github.com/iptoux/virtualMe.git
cd virtualMe
bun install
cd service && bun install
cd ../dashboard && bun installEvery dashboard connection requires a shared secret. Generate one before editing .env:
cd service
bun scripts/generate-secret.tsOutput:
Generated SERVICE_SECRET:
SERVICE_SECRET=4b9a1f3e8c2d7a6b...
Add this line to your service/.env file.
Keep this value — you will enter it in the dashboard on first launch.
cp service/.env.example service/.envEdit service/.env:
# X API credentials (OAuth 1.0a)
X_API_KEY=
X_API_SECRET=
X_ACCESS_TOKEN=
X_ACCESS_SECRET=
X_USER_ID= # Your numeric X user ID
# AI provider
AI_BASE_URL=http://localhost:1234/v1
AI_MODEL=local-model
AI_API_KEY=not-needed
# Service
SERVICE_PORT=3000
SERVICE_HOST=0.0.0.0
SERVICE_SECRET= # Required — generate with: openssl rand -hex 32All other settings (post frequency, active hours, persona, etc.) can be configured at runtime via the dashboard.
# Terminal 1 — service
cd service && bun run src/index.ts
# Terminal 2 — dashboard
cd dashboard && bun tauri devOn first dashboard launch, a Connection Settings dialog appears. Enter the service URL (http://localhost:3000) and the SERVICE_SECRET you generated. The dashboard stores both in localStorage — subsequent launches connect automatically.
cd service
bun run src/index.ts # production
bun --watch src/index.ts # dev mode with auto-restartThe service creates bot.db (SQLite) and bot.lock on first run. If it crashes without cleanup, delete bot.lock before restarting.
cd dashboard
bun tauri dev # development (hot reload)
bun tauri build # production buildOn first launch, a setup dialog asks for the service URL and secret — both are stored in localStorage and never transmitted anywhere else. The pencil icon in the sidebar lets you update these settings at any time.
All values below are startup defaults set via service/.env. Every setting is also editable at runtime via the dashboard's Config tab.
| Variable | Default | Description |
|---|---|---|
POLL_INTERVAL_MINUTES |
480 |
How often to read the home timeline (8h keeps within free tier) |
POST_FREQUENCY_HOURS |
2.5 |
How often to generate an original post |
REPLY_PROBABILITY |
0.4 |
Probability (0–1) of actually replying when the reply cycle runs |
MAX_POSTS_PER_DAY |
10 |
Daily cap on original posts |
MAX_REPLIES_PER_DAY |
6 |
Daily cap on replies |
ACTIVE_HOURS_START |
8 |
Posts and replies only happen between these hours (local time) |
ACTIVE_HOURS_END |
23 |
|
MAX_TWEET_LENGTH |
280 |
280 standard · 4000 X Premium · 25000 X Premium+ |
ENABLE_THREADS |
true |
Split posts longer than MAX_TWEET_LENGTH into reply chains |
PERSONA_PROMPT |
(see below) | The bot's voice and persona, sent as system prompt to the AI |
SERVICE_SECRET |
— | Required. Shared secret — the dashboard must provide this to connect. Generate with openssl rand -hex 32 |
The bot ships with an opinionated tech persona: blunt, direct, critical of PR nonsense and hype. Posts in the same language as the source article (German or English). Criticises ideas and decisions — never insults people personally. The full prompt is in service/src/config.ts and editable at runtime via the dashboard.
| Tab | What it shows |
|---|---|
| Overview | Uptime, posts/replies today, monthly API budget, manual trigger buttons |
| Live Logs | Real-time structured log stream with expandable metadata |
| Posts | Full post history with status badges; click ▶ to expand the AI prompt used to generate each post |
| News Sources | Add/remove/toggle RSS feeds and web URLs |
| Config | Edit all runtime settings including the persona prompt |
Real-time log stream from the service. Click the chevron on any entry to expand metadata.
Full post history. Click ▶ on any entry to reveal the exact system + user prompt the AI received when generating it.
Add and manage RSS feeds and web URLs. Toggle individual sources on/off without deleting them.
Service
- Bun — runtime, SQLite (
bun:sqlite), HTTP server - openai — AI client (OpenAI-compatible)
- rss-parser — RSS feed fetching
- OAuth 1.0a signing — implemented manually with built-in
crypto(no external Twitter SDK)
Dashboard
- Tauri v2 — native desktop shell (Rust)
- React 19 + TypeScript
- Tailwind CSS v4
- Radix UI primitives + shadcn-compatible components
- Vite — frontend bundler
virtualMe/
├── service/
│ └── src/
│ ├── index.ts # Entry point — Bun.serve()
│ ├── scheduler.ts # 60s tick loop, task registry
│ ├── bot-engine.ts # runPostCycle / runReplyCycle / runReadCycle
│ ├── ai-client.ts # generatePost / generateReply (returns text + stored prompt)
│ ├── x-client.ts # OAuth 1.0a, createPost, createReply, getHomeTimeline
│ ├── news.ts # RSS + URL fetching, content dedup
│ ├── search.ts # DuckDuckGo context enrichment
│ ├── db.ts # SQLite schema + query helpers
│ ├── config.ts # Two-layer config (env defaults + DB overrides)
│ ├── routes.ts # REST API handlers
│ ├── ws.ts # WebSocket server
│ └── logger.ts # Structured logger — DB + WebSocket broadcast
├── dashboard/
│ └── src/
│ ├── App.tsx # Sidebar nav + tab routing
│ ├── lib/
│ │ ├── api-client.ts # Typed REST wrapper + auto-reconnecting WsClient
│ │ └── store.ts # React hooks (useStats, useLogs, usePosts, …)
│ └── components/ # Feature components + shadcn-style UI primitives
├── shared/
│ └── types.ts # Shared TypeScript interfaces (Post, LogEntry, BotConfig, …)
└── service/.env.example
Service won't start — "lock file exists"
Delete service/bot.lock. It's left behind when the process is killed without a clean shutdown.
Replies never fire
The reply cycle only triggers every 4 hours, and the default REPLY_PROBABILITY of 0.4 means 60% of those attempts are skipped. Also check that the timeline cache is populated — use the Read Timeline button in the dashboard first.
AI returns empty content
If finish_reason is length in the logs, increase max_tokens or switch to a non-reasoning model. Reasoning models (DeepSeek-R1, etc.) consume extra tokens for their chain-of-thought.
Windows: dashboard won't start
Make sure you're using bun tauri dev from inside the dashboard/ directory. A root-level bunfig.toml can break Bun's bin resolution — the Tauri config uses node_modules\.bin\vite.exe directly as a workaround.
MIT