← all repositories
vellum-ai/vellum-assistant

An AI assistant that nags you on purpose

Vellum Assistant is a self-hostable personal agent designed to remember, reflect, and interrupt—politely—across Slack, Telegram, and your desktop.

917 stars TypeScript AgentsChat Assistants
vellum-assistant
Velocity · 7d
+11
★ / day
Trend
accelerating
star history

What it does

Vellum Assistant is a cross-platform personal AI that runs persistently and talks to you through macOS, iOS, web, Slack, Telegram, or Twilio. It stores structured memory—identity, preferences, projects, events—with local embeddings by default, then uses that memory to initiate conversations without being prompted. A trust engine resolves every actor as guardian, trusted, or unknown, and defaults to no access.

The interesting bit

The personality layer is genuinely odd: the assistant observes how you write during onboarding and generates its own SOUL.md and reflection journal. An hourly self-check loop lets it re-read notes, spot deadlines, and message you—unless you’re already mid-conversation. It’s the rare open-source project that treats proactivity as a feature, not a bug.

Key highlights

  • Eight memory types with per-type staleness windows, source attribution, and deduplication; hybrid dense + sparse retrieval
  • Model-agnostic: Claude, OpenAI, Gemini, or Ollama; embeddings default to local ONNX with cloud fallback
  • Sandboxed skills via manifest-driven plugins (SKILL.md + TOOLS.json) that inject tools at runtime
  • Fail-closed security: credentials isolated in a separate process, never exposed to the model; every tool sandboxed
  • Self-hostable local mode or managed cloud sign-in; the company behind it sells a separate commercial platform

Caveats

  • The CLI exists but the README calls it secondary: “the desktop app is our primary focus”
  • 638 stars suggests early traction; some ambitious features (proactive reasoning, self-authored personality) are hard to verify without extended use
  • iOS and web clients are mentioned in the description but the README emphasizes macOS; exact platform parity is unclear

Verdict

Worth a look if you want an assistant that actually owns its own state rather than waiting for prompts. Skip it if you need a mature, battle-tested automation layer—this is experimental infrastructure with strong opinions about identity and memory.

Frequently asked

What is vellum-ai/vellum-assistant?
Vellum Assistant is a self-hostable personal agent designed to remember, reflect, and interrupt—politely—across Slack, Telegram, and your desktop.
Is vellum-assistant open source?
Yes — vellum-ai/vellum-assistant is open source, released under the MIT license.
What language is vellum-assistant written in?
vellum-ai/vellum-assistant is primarily written in TypeScript.
How popular is vellum-assistant?
vellum-ai/vellum-assistant has 917 stars on GitHub and is currently accelerating.
Where can I find vellum-assistant?
vellum-ai/vellum-assistant is on GitHub at https://github.com/vellum-ai/vellum-assistant.

heatdrop uses Google Analytics to see which pages get read — nothing else. Your call. How we handle data.