Vellum assistant shutdown drives search for nine alternatives

Vellum has begun turning off its cloud-hosted assistants, leaving users to export their data and find another platform. ZenML reviews nine alternatives for people replacing the assistant and teams that used Vellum to build, evaluate, and debug LLM applications.
- Vellum began switching off cloud-hosted assistants on November 7, 2026.
- Assistant exports, downloads, and logins close on December 13, 2026.
- ZenML reviewed nine alternatives for assistant users and LLM development teams.
- Kitaru replays agent sessions using recorded tool results and supports CI evaluation gates.
- The Vellum assistant repository remains available under the MIT license but is not actively maintained by Vellum.
Vellum users face an export deadline
Vellum began taking final exports and switching off cloud-hosted assistants on November 7, 2026, according to ZenML. Users can sign in at vellum.ai/export and download an assistant’s .vbundle file until December 13, 2026. After that date, downloads and logins will close, and Vellum will permanently delete the hosted data.
ZenML says Vellum’s restore instructions explain how to run an assistant locally, but saved credentials are not included in exports, so services must be connected again. Assistants already running locally with users’ own model keys and without a Vellum account can continue working. The assistant repository remains available under the MIT license, although Vellum will not actively maintain it. A founding engineer is maintaining Cadet, a rebranded fork for users who want to stay close to the original assistant.
The shutdown notice does not say that Vellum’s older development services have also ended. Those services covered prompt management, workflow building, deployments, and evaluations. However, ZenML says the development changelog has no entry after February 2026, so teams are advised to check support status and export their work.
Nine alternatives cover different migration needs
ZenML evaluated alternatives across migration and data portability, agent evaluation, production replay and reproducibility, and observability and debugging. The review covers both groups affected by the change: assistant users and teams that built applications on Vellum.
- OpenClaw is presented as an option for replacing the Vellum assistant, with messaging channels, local memory, and skills. It is free and open source, while model costs are extra.
- Dify targets teams rebuilding Vellum workflows in a visual builder, with Workflow Studio, agents, RAG pipelines, and node-by-node run logs.
- Langflow is a free, MIT-licensed visual builder whose flows can be stored as portable JSON.
- LangSmith is aimed at LangGraph teams needing traces and evaluations, including LLM judges, pairwise and human evaluations, and OTLP ingest.
- Braintrust focuses on evaluation-first teams that gate merges on scores, with scorers, human review, and CI support.
- Langfuse, Comet Opik, and Arize Phoenix focus on trace collection, test suites, and observability, with different self-hosted and hosted options.
- Kitaru by ZenML Labs focuses on replaying production sessions against a controlled change.
Kitaru focuses on replaying real agent sessions
ZenML describes Kitaru as an open-source, replay-based evaluation runtime for AI agents. It takes an existing session, changes one element, reruns the agent’s real code, and scores the new result next to the original. It does not replace Vellum’s workflow builder or assistant, but is intended for the evaluation stage after migration.
Kitaru can import complete sessions from Langfuse, LangSmith, Braintrust, Logfire, Arize Phoenix, MLflow, and Kitaru JSONL. There is no Vellum importer, although ZenML says users can write a custom importer. During replay, recorded tool results can be returned instead of making live calls. Users can also apply a model, prompt, or code change and compare the result with the unchanged baseline.
The platform includes versioned Python evaluators, LLM judges, custom rules, deterministic evaluator bundles, expert review, and a CI gate that can stop a change when a run fails. Kitaru is free to self-host under Apache 2.0. Its managed Pro Starter option costs $39 per month and includes one AI agent, one ML project, and up to 100 pipeline runs per month. A Pro Enterprise option is custom priced and supports on-premises or air-gapped deployment.
For users, the immediate priority is exporting hosted assistants before the December 13 deadline. For development teams, ZenML’s review suggests that the right replacement depends on whether they need a visual workflow builder, an assistant, tracing and evaluation, or controlled replay of production failures.
ZenML: 9 Vellum Alternatives After the Assistant Shutdown


