AI Briefing
KOSign in

Traceway: Self-Hosted Observability Platform Integrating LLM Cost Tracking and AI Agent Queries

·2026.10.20 11:30

Key point

Traceway, an MIT-licensed self-hosted observability platform, supports LLM cost tracking and AI agent integration.

1 / 4

Details

  • Traceway, a self-hosted observability platform based on OpenTelemetry, integrates logs, traces, metrics, exceptions, session replay, and LLM call tracking into a single server.
  • LLM calls are organized as spans with gen_ai.* attributes by model, token, cost, conversation, and tool call units; cost calculation requires either sending values from external services like OpenRouter or direct calculation within the app.
  • Coding agents (such as Claude Code, Cursor, and Codex) can query operational data via skills, a read-only CLI, and MCP servers, while write operations require separate credentials and organization administrator approval.
  • Self-hosting supports configurations with SQLite, DuckDB, and ClickHouse + PostgreSQL; version 2.0 is a breaking upgrade due to changes in the trace table structure.
  • Traceway Cloud offers a free plan and a Premium plan at $24.99 per month ($0.25 per GB for overage), while self-hosting is free with no feature differences.
  • Compared to commercial APMs and open-source stacks, its strengths include the MIT license, installation via docker compose up -d, direct OTLP/HTTP ingestion, and integrated replay/trace/AI call tracking.
  • Due to the absence of a gRPC receiver, gRPC-default SDKs such as the Python SDK require the OTEL_EXPORTER_OTLP_PROTOCOL=http/protobuf setting.
  • The DUCKDB_MEMORY_LIMIT setting is mandatory for DuckDB configurations; failure to set it may result in OOM errors.
  • The system recorded response times under 5 seconds when querying 100 million data points on a 2 vCPU/8GB machine, while SQLite reaches its limit near 1 million points.
  • Agent integration installs 4 skills via the npx skills add tracewayapp/traceway --all command, and Claude Code is the only agent directly tested by the development team.
observabilityai-agentsself-hostedopentelemetryllm-monitoringcost-trackingtraceway
Read originalShare on Threads

This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.

Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.

Related

Claude BlogClaude Platform Adds Native OpenTelemetry Export for Agent ObservabilityHacker NewsLaminar releases flow-1, an RL-trained model for detecting silent agent failuresNetflix TechNetflix Builds Ontology-Driven Knowledge Graph to Streamline Incident Context AssemblyLangChain BlogLangSmith Team Answers Common Questions on Agent Evals, Highlighting Cost Reductions and Framework AgnosticismNVIDIA BlogNVIDIA Launches RTX Spark PC for Creators, Gamers, and AI AgentsNVIDIA BlogNVIDIA Releases Nemotron 3.5 Super VL and Bolt Models for Edge and Enterprise Agents