OpenLIT

Trace full agent workflows instead of treating each LLM call as an isolated event

OpenLIT represents prompts/responses, tool calls, retrieval, embeddings, memory, sub-agents, token usage, cost, errors, and evaluations as OpenTelemetry telemetry, making root-cause investigation possible across the complete workflow.

Sources: [2]

Instrument Claude Code, Cursor, and Codex sessions from the CLI

Its coding-agent instrumentation can capture prompts, LLM calls, tools, file reads and edits, shell commands, searches, sub-agent activity, token usage, cost, and code impact from supported agent sessions.

Sources: [2]

Version 2.1.0 adds stuck-loop detection and trace governance

OpenLIT 2.1.0 detects repeated tool calls as stuck-agent loops and reports wasted tokens and cost. It also adds trace governance, coding-agent security checks, policy packs for NIST AI RMF, the EU AI Act, and OWASP ASI, plus Governance Passport export.

Sources: [3]

For teams operating AI or coding agents that need tool-level debugging, cost analysis, and quality evaluation

It is a strong fit when an agent spans multiple providers and frameworks and teams need to identify which prompt, tool, retrieval, or sub-agent caused a failure while keeping telemetry compatible with OpenTelemetry infrastructure.

Treat agent telemetry as sensitive data

Coding-agent observability can capture prompts, LLM traffic, file operations, and shell commands. Assume source code, secrets, or personal data may enter telemetry and design export destinations, retention, access controls, and redaction accordingly.

Sources: [2]

The default Compose stack exposes ClickHouse and OTLP ports and includes fallback database credentials

The repository Compose file exposes ClickHouse on 9000/8123, OpenLIT on 3000, and OTLP on 4317/4318, with default database user/password values. Change credentials and design network exposure, TLS/authentication, and volume backups before shared or production use.

Sources: [4][3]