AI Potluck
Product / UX / Orchestration & agents

Vellum

Vellum

Personal AI assistant across web, macOS, iOS and a CLI, with episodic, semantic and procedural memory, that takes actions on the user's behalf across channels including Slack and Telegram. It runs proactive check-ins and isolates credentials out of process. The company pivoted to it in 2026 from an LLMops platform, which remains available.

Verified 2026-08-13 via the vellum-ai/vellum-assistant repository, its LICENSE and the Vellum product site.

Openness

5 high confidence
5.0
license
MIT(OSI, Copyright 2025 Vellum AI)
source
public(the full agent - gateway, clients, skill framework, memory - in github.com/vellum-ai/vellum-assistant)
core-gated
ungated(README: managed runtime on Vellum Platform or self-hosted, "Same codebase, same data model")
self-host
yes(by design)
managed-cloud
hosting convenience, not a feature tier

The agent's complete source is public and MIT-licensed (the LICENSE reads "Copyright (c) 2025 Vellum AI"), covering the backend gateway, the Mac, Slack and Telegram clients, the sandboxed skill framework and the memory system, and it is self-hostable by design rather than a client stub. The vendor's own marketing comparison table labels Vellum "Proprietary" against MIT for Hermes and OpenClaw, because it sells a managed cloud tier - but the same comparison post says outright that "Vellum is also open source", so that label is not a live caveat. The managed memory and security layer is not a feature tier either: the README states that the managed runtime and the self-hosted one are the "Same codebase, same data model", so hosting is a convenience rather than a gate, and a hosted offering sold alongside a complete open core is not gating.

Adoption

2 low confidence
2.0

1,054 stars on vellum-ai/vellum-assistant, which bands at 1K-10K stars, level 2 on the stars scale. No download, user or usage metric is disclosed - this is a self-hosted app rather than a registry package - so stars are the only signal available, and they are used strictly as a last-resort proxy. A small, new entrant.

Capability

3 low confidence
3.0

Solid personal-agent feature breadth, all of it visible in the open MIT source: 8-type memory, real-world action through managed OAuth connections, multi-channel reach across Mac, web, iOS, Slack, Telegram and phone, out-of-process credential isolation and an ambient proactive engine. No benchmark and no frontier differentiator, so this is a mid-tier judgment on features.

  • https://www.vellum.ai recorded 2026-08-13

    structured memory, multi-channel autonomous action, credential isolation and an ambient/proactive engine; no benchmark figure published

Verified 2026-08-13