AgentTrace

Local-first TUI and reports for AI coding-agent session history, cost, tokens, time, and slow-run diagnosis.

agenttrace running locally against real AI coding agent session logs

--- **agenttrace** is a local-first terminal TUI and report generator for AI coding-agent session history. It reads Claude Code, Codex CLI, Gemini CLI, Qwen Code, Cline, Aider, Cursor exports, Hermes Agent, OpenCode, OpenClaw, Pi, Oh My Pi, Kimi CLI, Copilot-style logs, and generic JSON/JSONL traces, then helps with two daily jobs: see what multiple agents spent across cost, tokens, and time; and diagnose why a task ran slowly. One `agenttrace` binary provides both interfaces: run it without a report action to open the TUI, or pass flags such as `--sessions` and `--overview` for CLI output. ## Why agenttrace? AI coding agents now behave like small build systems: they call tools, retry, stall, and spend tokens while you only see the final answer. **agenttrace** reads the logs your agents already write and puts cost-heavy or slow sessions first. It helps you answer: - **What did my agents spend?** Compare historical sessions by agent source, model, input/output/cache tokens, estimated cost, and wall-clock time. - **Why was this task slow?** Catch long gaps, hanging sessions, retry loops, slow tool calls, large parameters, and context pressure. - **Did a run regress?** Compare against a local baseline when supplied, then inspect incident timelines and conservative tool authority categories in reports. - **What should I inspect first?** Rank sessions by cost, duration, turns, health, failures, anomalies, model, source, or text search. - **Can I inspect this privately?** Everything runs locally; prompts, code, and logs do not need to leave your machine. ## Real local run These screenshots and figures are a redacted sample of the latest 500 real local sessions captured for the v0.6.0 source tree. They are not `--demo` output, current telemetry, or test fixtures. ```bash agenttrace ``` | Overview | Critical sessions | |---|---| | agenttrace overview showing real local AI coding agent sessions, token cost, errors, and health

agenttrace overview showing real local AI coding agent sessions, token cost, errors, and health

agenttrace critical session list from real local AI coding agent logs

| | Session detail | Diagnostics | |---|---| | agenttrace detail view showing health, cost, tool failures, and next action from a real local session

agenttrace diagnostics view showing latency, context window, and large parameter calls from real local logs

| That local run found: ```text AGENTTRACE v0.6.0 ``` | Signal | What agenttrace found | |---|---:| | Analyzed sessions | 500 | | Total tokens | 4.0B | | Estimated cost | $2.6K | | Tool failure rate | 2.9% | | Critical sessions | 2 | | Average health | 77.7% | ## Install The commands below install the latest public channel, which can lag behind the `v0.6.0` source tree until its release assets and tap formula are published. Check the installed version with `agenttrace --version`. To test the current source tree, use the Git-based Cargo command below. ```bash curl -sL https://raw.githubusercontent.com/luoyuctl/agenttrace/master/install.sh | sh ``` Other install paths: ```bash brew install luoyuctl/tap/agenttrace cargo install --git https://github.com/luoyuctl/agenttrace agenttrace ``` Windows: ```powershell iwr -useb https://raw.githubusercontent.com/luoyuctl/agenttrace/master/install.ps1 | iex ``` ## Quickstart ```bash agenttrace ``` ## What you get | Need | agenttrace gives you | |---|---| | Historical spend review | Sessions grouped across projects, agents, and models with Today/7d/30d/All ranges | | Data confidence | Parse skips, cache hits, unknown sources/models, pricing fallbacks, and latest observed session | | Honest capability levels | `Detailed`, `Aggregate`, or `Limited` per session so missing event-level evidence is never presented as a complete trace | | Privacy-safe steps | Tool-step metadata and duration when the source provides call IDs and timestamps; no prompt, response, result, or tool-argument body is stored in steps | | Slow-task diagnosis | Latency stats, long gaps, hanging sessions, retry loops, slow tools, large params, and context pressure | | Regression evidence | Local baseline comparison when supplied, incident timelines, and conservative tool authority categories in reports | | First-session triage | Sort and filter by cost, duration, health, failures, anomalies, model, source, or text search | | Shareable evidence | JSON, Markdown, and self-contained HTML reports | | Local-first inspection | No hosted backend required | ## Docs - Site: https://luoyuctl.github.io/agenttrace/ - Sample HTML report: https://luoyuctl.github.io/agenttrace/demo-report.html - CI setup: [docs/ci-integration.md](docs/ci-integration.md) - Cursor import: [docs/cursor-import.md](docs/cursor-import.md) - Parser guide: [docs/parser-guide.md](docs/parser-guide.md) - Launch notes: [docs/launch-kit.md](docs/launch-kit.md) Listed in these open source projects: - [awesome-mac](https://github.com/jaywcjlove/awesome-mac) - [antigravity-awesome-skills](https://github.com/sickn33/antigravity-awesome-skills) - [awesome-claude-skills](https://github.com/BehiSecc/awesome-claude-skills) ## Contributing Parser PRs are welcome. A good parser contribution usually includes: - a tiny redacted fixture or synthetic sample - format detection in `crates/agenttrace-core/src/parser.rs` - role, timestamp, model, token usage, tool call, and tool error extraction - tests for successful parsing and malformed input Generated parser fixtures are deterministic: ```bash python3 scripts/generate-testdata.py python3 scripts/generate-testdata.py --check ``` Run before sending a PR: ```bash cargo test python3 scripts/generate-testdata.py --check cargo build --release -p agenttrace target/release/agenttrace --doctor ``` See [CONTRIBUTING.md](CONTRIBUTING.md) for the full contribution flow. ## License [MIT](LICENSE) © 2026 agenttrace contributors