← Back to changelog

August 21, 2026

Langfuse CLI 1.0

Nimar Blume

The CLI is rewritten from the ground up. Invocations start 10x+ faster, union endpoints take plain flags, and failures exit with a code agents can read.

The Langfuse CLI is at 1.0 with a rewritten engine. Every invocation starts 10x+ faster, union endpoints take plain flags instead of hand-written JSON, and agents can tell what went wrong from the exit code alone.

A few things that get easier:

# Create a prompt in one line, no JSON body to assemble
langfuse api prompts create --type text --name greeting --prompt 'Hello {{name}}'

# Export a full result set instead of one page at a time
langfuse api observations list --limit 100 --all --max-items 5000

# Point a script at an older self-hosted deployment
langfuse --api-version auto api prompts list

Install or upgrade:

# upgrade
npm install -g langfuse-cli@latest
# upgrade via bun
bun add --global langfuse-cli@latest

# or run directly via npx/bunx
npx langfuse-cli api <resource> <action>
bunx langfuse-cli api <resource> <action>

What's new:

• 10x+ faster startup time on every invocation.
• Typed flags for union endpoints (prompts, evaluators, placements) with local validation, so the discriminator flag selects the variant and that variant's own required fields apply.
• --all pagination walks every page of a list, page-based or cursor-based, bounded by --max-items and a request cap.
• Consistent flags, with every parameter name in kebab-case (--object-id). --body-json and --body-file - remain the lossless path for complex nested bodies.
• Zero runtime dependencies, running on Node.js 20+ or Bun.
• Machine-readable exit codes so agents classify failures without parsing stderr: usage (2), configuration (3), network (4), HTTP failure (5), and local errors (6).
• Works with every Langfuse Server since v3. Pin an API snapshot with --api-version 3, or let --api-version auto detect your server and select a compatible one. Defaults to the latest.

Upgrading from 0.x: body-field flags changed from camelCase to kebab-case (--objectId becomes --object-id) and HTTP failures now exit 5 instead of 1. --body-json inputs are unaffected. Details in the GitHub repo.