AgentMon Live – Watch Claude Code agents and check their work – Apache 2.0

Codenotary launched AgentMon Live, an open-source monitoring and verification plugin published under the Apache 2.0 license for Anthropic's Claude Code command-line tool. Designed to make autonomous coding sessions transparent, the plugin runs directly in the terminal or through a dedicated browser interface, streaming subagent execution paths, tool calls, costs, and token consumption without sending session data to remote servers. Software developers can examine the project's source code and repository documentation through github.com's report detailing the release.

This breakdown covers how AgentMon Live instruments Claude Code sessions, intercepts autonomous workflows, and checks whether code changes actually pass testing thresholds before an agent reports completion. The tooling is targeted at software engineers, technical leads, and development teams running Claude Code 2.1.308 or later who need visibility into nested subagent lifecycles, rate limits, and context windows, while preventing premature handoffs caused by unverified fixes or broken builds.

What shipped in AgentMon Live – Watch Claude Code agents

AgentMon Live introduces a multi-section local monitoring interface accessible directly beside terminal sessions or as a web dashboard served on 127.0.0.1. The interface splits into an Overview rail, Tool calls, Verification, Subagents, Files, and a Session log. Terminal users gain interactive hotkeys to inspect slow operations, expand collapsed subagent pools past four items, mask secrets automatically, and toggle between calm and classic pane views.

The browser experience includes an agent flow visualizer that tracks ancestry trees to show which parent processes spawned specific subagents. An integrated waterfall view nests calls under executing agents, mapping elapsed turn times against active waiting approvals. The verification engine checks the status of the latest test, type check, linter, and build runs, explicitly highlighting files that changed after the last successful check run.

What improved in AgentMon Live – Watch Claude Code agents

Earlier monitoring setups lacked native safeguards against silent test failures or circular execution loops, often requiring developers to scan verbose command-line outputs manually. AgentMon Live embeds an autonomous done gate that intercepts completion claims, running verification tests directly and redirecting Claude Code back to debugging if test suites fail. The system also tracks tool performance across sessions, identifying flaky utilities and detecting drift inside project configuration files like CLAUDE.md.

Operational awareness improves via proactive alerts that trigger when subagents stall or cycle through repeated errors. Developers can run comparative benchmarks to measure session efficiency with and without monitoring active.

What you gain from AgentMon Live – Watch Claude Code agents

Engineers gain immediate financial and resource oversight through meters that track context window consumption, quota exhaustion, and cumulative session costs. Visual pins move dynamically to flag when cleaning context or compacting conversations would prevent wasted spend. When an agent gets blocked on duplicate approvals or recursive execution errors, pressing a dedicated key drafts an immediate follow-up query to redirect the agent.

The software also protects repository health by sorting edited files by test freshness, flagging files that were touched after the last passing test run. Subagent tables sort by error rates, duration, and token usage to expose runaway processes quickly. The entire plugin operates locally in six interface languages, adapting to English, German, French, Spanish, Chinese, or Japanese based on the host terminal environment.

How to get AgentMon Live – Watch Claude Code agents

Installation requires Claude Code version 2.1.308 or later and a Node.js binary present on the system path to serve the local browser interface. Users can install the plugin inside an existing Claude Code session by pointing to the official marketplace repository:

/plugin install agentmon-live --marketplace codenotary/agentmon-live

Developers managing plugins from an external command prompt or scripting unattended machine setups can invoke the standard command-line syntax:

claude plugin install agentmon-live --marketplace codenotary/agentmon-live

To configure secondary model evaluations, users can launch sessions with specific models pinned or modify their local settings file:

claude --model fable

Inside active sessions, operators switch model advisors by issuing slash commands:

/model claude-fable-5-1

Default behavior can be locked without model calls by setting advisorMode to gate in the local plugin store settings.

What to watch after AgentMon Live – Watch Claude Code agents

Security posture requires consideration when working across multi-user environments. Because the web dashboard binds to localhost port 127.0.0.1 without authentication, other users sharing the same physical or virtual machine can view running sessions unless the web interface is explicitly disabled in settings. Secrets are masked before display, but teams on shared jump hosts must disable the local port to enforce isolation.

Developers should also monitor optional background model checks handled by Fable. When advisor mode is enabled, Fable consumes roughly 12,000 characters of session text to verify that every user constraint received an answer, capped at 30 checks per session by default. Teams prioritizing strict token conservation should verify that the done gate runs in local-only gate mode to eliminate extra model calls while retaining local test verification.