**Note:** AgentLens is now **TraceRoost**. Already using AgentLens? See [Upgrading from AgentLens](#upgrading-from-agentlens).

[](https://github.com/traceroost/core/actions/workflows/ci.yml) [](LICENSE)


Local monitoring and observability for agentic AI coding tools — see what's actually happening inside each run. Nothing leaves your machine.

TraceRoost receives **OpenTelemetry traces** from Copilot, Claude Code, and Codex in real time, giving you span timing, time-to-first-token, per-tool latency, and file diffs. It also reads the **local log files** each agent writes automatically — including OpenCode's **SQLite database** — as a zero-config fallback that backfills history from before you set anything up. Both sources appear in one dashboard; OTEL takes precedence when available.

Two things it does that a usage dashboard doesn't:

• Catches agents that are stuck. Ten named loop and malfunction patterns — repeated tool calls, oscillating edits, recurring errors, runaway scope, hallucinated dependencies, unverified test runs, and more — each with a correction prompt you can paste straight into the session. See the full list →
• Tells you what to fix in your instructions file. The Advisor reads across traces and suggests concrete additions to your CLAUDE.md or AGENTS.md — including hot files the agent rediscovers from scratch on every run. More →

**Quick start:**

npx traceroost@latest

Open — that's it. Running it in a terminal only lasts until you close it, though: if TraceRoost isn't running when an agent sends OTEL data, that