agtrace

See What Your AI Agent Is Actually Doing


---

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.

!agtrace with Claude Code

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.

---

!agtrace watch demo

Try It

``bash
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:

`bash


Claude Code

`bash
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:

`toml
[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