Codenotary today released AgentMon Live, a free, open-source plugin that gives Claude Code users real-time visibility into what their agents are actually doing. Available now at github.com/codenotary/agentmon-live under Apache 2.0, it installs in a single command and requires Claude Code 2.1.308 or later.
The problem it solves has been growing quietly alongside Claude Code’s adoption: when you hand a task to an agent, you largely have to take its word for what happened. No visibility into which tools ran, how tokens were spent, whether subagents finished, or whether “done” actually means done.
The Done Gate: Agents Must Prove Completion
The most important feature is the Done Gate. It examines recorded session activity for unresolved failures and checks whether every edited source file has had a corresponding test run since the last edit — at the package level, not globally. If a package was touched but its tests haven’t run since, Done Gate prompts Claude Code to finish the work or explicitly explain what remains.
This matters because the common failure mode of AI coding agents isn’t catastrophic errors — it’s silent non-completion. An agent declares a task done, you move on, and the work is half-finished. The Done Gate introduces a verifiable completion standard rather than a self-reported one. Dennis Zimmer, Codenotary’s CTO and co-founder, put it plainly: “AI coding productivity is measured by how much useful, verified work gets completed.”
Fable: A Second Model Watching Your Agent
AgentMon Live runs a second model alongside your session called Fable. It reads agent activity in real time and speaks up only when something looks off: a plan that skips something obvious, three consecutive errors, or a task declared complete with parts still missing. It doesn’t narrate every step — it only flags anomalies.
The concept is essentially a code reviewer for agent sessions. Using one model to monitor another is not common in the Claude Code ecosystem, and the approach is worth watching. Fable won’t catch everything, but it surfaces the class of failures that are easy to miss when you’re not watching the session live.
Browser View and Token Visibility
A recent addition is the browser view, which lets you follow agent activity from a web browser rather than staring at a terminal pane. It shows agent flow (which agent spawned which), a waterfall of every tool call, and a top band with context usage, quota consumption, cache hit rate, cost, and turn count. If you’re running parallel sessions across multiple terminals, this gives you a single view without context switching.
The token breakdown is practical. AgentMon Live surfaces which MCP servers you loaded but never called in a session, and tells you when /clear or /compact would actually reduce costs. Most developers have no sense of which part of their workflow is expensive. This makes the invisible visible without any configuration.
How to Install
Installation is one command. Run this inside a Claude Code session:
/plugin install agentmon-live --marketplace codenotary/agentmon-liveOr from the shell:
claude plugin install agentmon-live --marketplace codenotary/agentmon-liveThe plugin appears as a live pane beside the chat. No separate configuration required. Full details are in the official Codenotary blog post. The creator’s own write-up on DEV Community covers the browser view addition in detail. For a comparison with other harnesses, see SD Times’ coverage.
The Bigger Picture
Agentic workflows are maturing quickly but unevenly. The models are capable. The tooling for understanding what they’re doing is still catching up. AgentMon Live is an early but concrete step toward treating Claude Code sessions the way you’d treat any other piece of infrastructure — with logs, traces, and something that tells you when a job actually finished.
It’s version 0.36.1 and actively developed. Some rough edges should be expected. But the Done Gate alone is worth installing for anyone running Claude Code on tasks that matter.
