NEWZero PolarDB-X is now in public beta. Instant database, no sign-up required.
langfuse SDK compatible

See every LLM call

PXFuse records every reasoning step, every token and every cent your agent spends. langfuse SDK compatible β€” three lines of config to connect.

Supported Agent Types

⚑ CodexπŸ€– Claude CodeπŸ”· Qoder🏒 QoderWorkπŸͺ½ Hermes

Read https://zero.polardbx.com/pxfuse/SKILL.md and follow the instructions to set up observability for your LLM app with your PXFuse project.

This is what one agent call looks like in PXFuse

trace / a3f1c8e2-7b04-4d19
total 1,240ms1,589 tokenscost $0.0125
agent.run1240ms
β”” retrieve.docs180ms
β”” llm.chat620ms
β”” tool.weather150ms
β”” llm.summarise210ms
gpt-4o Β· 1,203 tok Β· $0.0121gpt-4o-mini Β· 386 tok Β· $0.0004

Three lines to connect

PXFuse speaks the langfuse protocol β€” swap three environment variables and you are done

.env
LANGFUSE_BASEURL=https://pxfuse.example.com
LANGFUSE_HOST=https://pxfuse.example.com
LANGFUSE_PUBLIC_KEY=pk-lf-xxxxxxxx-xxxx-xxxx-xxxx-xxxxxxxxxxxx
LANGFUSE_SECRET_KEY=sk-lf-xxxxxxxx-xxxx-xxxx-xxxx-xxxxxxxxxxxx

Sample values below. Claim a project and your real credentials appear here automatically.

My Projects

Manage your provisioned PXFuse projects

Capability Map

Trace Β· Measure Β· Evaluate Β· Govern β€” four layers converging on production quality

πŸ“Š

Tracing

Tracing

Full-path traceability

  • Full-path call tracing
  • Nested span hierarchy
  • Session context linking
  • Error and retry annotation
πŸ“ˆ

Metrics

Metrics

Calls quantified as metrics

  • Throughput, latency, error rate
  • Token usage and cost accounting
  • Model, project, user drill-down
  • Custom metric ingestion
πŸ§ͺ

Evaluation

Evaluation

Output quality made measurable

  • Human annotation and scoring
  • LLM-as-judge scoring
  • Badcase-driven datasets
  • Cross-version regression
πŸ›‘οΈ

Governance

Governance

Changes controlled and reversible

  • Prompt version control
  • Staged rollout and rollback
  • Project-level data isolation
  • Per-project key authorization

From connect to operate

Four steps that turn LLM app quality into a loop you can keep iterating on

πŸ”Œ

Connect

Drop in the host and key pair β€” calls report automatically, no logic changes

πŸ”

Debug

Expand your agent's multi-step tool calls and see exactly where it stalled

⚑

Optimise

Surface the slow, expensive calls, then go back and tune prompts and models

πŸ“Š

Operate

Sample, score, regress badcases β€” let the data decide the next revision