Skip to content
Vellum → EvalGuard

Bring your executions. Get security with your evals. 

Vellum runs your prompt and workflow executions. EvalGuard imports that run history and adds red team (300+ plugins), a firewall on the gateway hop, and a BYOK gateway alongside prompt and eval tooling — one workspace, one bill.

SOC 2 evidence engineISO 42001 mappedEU AI ActGDPR

Prompts + evals were the start

What you keep — and what you gain

Keep
  • Prompt- + workflow-execution history
  • Model per execution (provider too, when your export carries one)
  • Inputs + outputs (variables and completion blocks)
  • Token usage — prompt / completion
  • Cost + latency per run
  • Execution state, and every other column you export
Gain
  • + Firewall on the gateway hop rather than a node inside a workflow — 439 DLP patterns, input and output
  • + 300+ red-team plugins + 100+ adversarial strategies
  • + 200+ eval scorers (LLM-as-judge, pairwise, rubric)
  • + Prompt IDE + optimizer across 90+ providers
  • + Managed BYOK gateway with similarity response caching
  • + 50 compliance frameworks mapped to live evidence + a tamper-evident audit log

Backfill your history

Bring your Vellum executions along

Export your executions from Vellum (the Executions tables in the deployment UI at app.vellum.ai, or their workflow- and prompt-deployment executions APIs), then convert them with the EvalGuard CLI. Both shapes are accepted, no record is dropped, and re-running the same import never double-counts, because spans use a stable content hash.

# Convert your Vellum export to neutral-shape spans:
npx @evalguard/cli import:traces --from vellum vellum-export.json --output spans.json

Prints an import summary (spans imported / duplicates skipped / parse errors) and writes the neutral spans to spans.json. This is an execution-history import — it moves your run logs, not your Vellum prompt deployments or workflow definitions.

What importsWhere it lands
Workflow + prompt executionsOne EvalGuard span each, grouped by trace id
Model (and provider, when present)span.model / span.provider
Inputs / outputsspan.input / span.output
Token usage (prompt / completion)span.promptTokens / completionTokens / totalTokens
Costspan.costUsd
Latencyspan.durationMs (latency, latency_ms, or start / end)
Execution state + every other columnspan.attributes (vellum.* namespace)

Vellum's terminal failure state is REJECTED; those executions import with status: error so your failure history survives the move. Everything else the importer recognizes is preserved verbatim under span.attributes.

Vellum claims on this page were checked against their published documentation docs.vellum.ai + the public vellum-ai SDK types (raw) — on 2026-08-10. Capabilities move; check their current docs before you decide.

Note that vellum.ai now fronts a consumer personal-AI assistant; the developer platform this page is about is at docs.vellum.ai and app.vellum.ai.

One platform, one bill

Ship prompts. Secure, eval, optimize — everywhere.

Prompt tooling is the hook. The platform is why you stay.

Start free