third3y3.com

What It Is

thirdeye is a local-first observability tool for AI coding agents, with the core positioning: “Trace every agent session, locally.” It does not provide a cloud SaaS service. Instead, it stores traces, tags, evaluation results, and token usage data on your own machine under ~/.thirdeye/, and lets you query and analyze them via a CLI and local Browser UI.

Core Capabilities

In terms of functionality, thirdeye covers the key parts of auditing AI agent sessions. You can use thirdeye add to register hooks for Claude Code, Cursor, OpenAI Codex CLI, Gemini CLI, and GitHub Copilot CLI. After that, you can inspect history with list, events, tail, and search, manage event tags with tag, and summarize token consumption with usage. It also supports rubric-based eval run, allowing claude, codex, and gemini to act as LLM-as-judge evaluators for sessions. On the data side, it uses JSONL sidecars and a SQLite usage index, and most read commands support --json, making it suitable for scripting and automation.

Pricing and Deployment

The main documentation does not disclose any paid plans or enterprise pricing. The product emphasizes No cloud, No SaaS, and No telemetry, and offers installation via Homebrew, pipx, and uv. It is naturally a local deployment tool, with the data directory configurable via THIRDEYE_HOME. The Browser UI runs by default on 127.0.0.1:8765 and requires installing the UI extra.

Pros and Cons

Its strengths are that it is privacy-friendly, requires no account, and can track supported agent CLIs without requiring changes to business code. The command system is detailed, and the documentation lists complete subcommands, parameters, and examples, making it easy to get started. The downsides are that the main documentation only explicitly mentions Mac and Linux, with no Windows support clarified; coverage depends on the agents already supported; the evaluation feature still requires external LLMs/agents to be available; and there is no visible information about team collaboration, permissions, shared dashboards, or commercial support.

Who It’s For and Access from China

It is well suited to individual developers and small teams that heavily use Claude Code, Cursor, Codex, Gemini, or Copilot CLI, especially for reviewing agent actions, troubleshooting unintended changes, tracking token usage, and building reusable evaluation rubrics. Access from China is not specified in the main documentation. Installation may involve Homebrew, PyPI, GitHub, and related agent services, so network access and model service availability should be verified independently. If you need cloud-based team observability, alternatives such as Langfuse, Helicone, and Arize Phoenix may be worth comparing.