The Problem I Had

When I started using AI coding agents (Claude Code, Codex, Gemini), I realized I was working with a black box. I couldn't see:

• How much of the context window was being consumed
• What the agent was actually doing between my prompts
• When the conversation was getting too long and performance would degrade

I found myself guessing the agent's internal state. That felt wrong.

What Changed

Now I always run agtrace alongside my coding agent. It's become essential.

What I see:

• Context window usage — A color-coded bar showing how full the conversation is
• Token consumption trends — How much context each task uses over time
• Live activity — Tool calls, file reads, reasoning traces as they happen

For the first time, I can make informed decisions about when to start a new session, how to scope my requests, and whether the agent is stuck in a loop.

Try It

npm install -g @lanegrid/agtrace
cd my-project
agtrace init      # One-time setup
agtrace watch     # Launch dashboard in a separate terminal

Works with Claude Code, Codex (OpenAI), and Gemini. Zero config — just discovers existing logs.

Give Your Agent Memory of Past Sessions

One thing I didn't expect: agents can also query their own execution history via MCP:

# Claude Code
claude mcp add agtrace -- agtrace mcp serve

# Codex
codex mcp add agtrace -- agtrace mcp serve

Now your agent can search what it did yesterday, find past errors, and learn from previous sessions.

See the MCP Integration Guide for more.

Other Commands

agtrace session list       # Browse past sessions
agtrace lab grep "error"   # Search across all sessions

For Tool Builders

If you're building your own IDE plugin, dashboard, or observability tool:

[dependencies]
agtrace-sdk = "0.6"

See SDK Documentation and Examples.

Documentation

• Getting Started
• MCP Integration
• Architecture
• Full Documentation

Feedback

Have ideas?

• RFC: Watch TUI Display
• RFC: MCP Tools

License

MIT / Apache 2.0