Codenotary's free AgentMon Live checks whether Claude Code finished the job

Agent activity during a session

Claude Code already prints a running log of its actions into the terminal. That log scrolls. Once an agent has read twelve files, patched three, run the test suite twice and summarised its work, the evidence sits above the current prompt, buried in output many users will not read line by line.

AgentMon Live adds a structured view on top of that stream. According to Codenotary's announcement, the dashboard reports active tools, subagents, file modifications, pending approvals and verification results. It also shows token consumption, context growth, cache efficiency and tool execution time, and it identifies potential inefficiencies including repeated file reads, excessively large tool outputs and unused Model Context Protocol servers. A built-in benchmarking command lets developers compare sessions with and without AgentMon Live.

The source code is published at github.com/codenotary/agentmon-live, so the logic can be inspected before anyone runs it against a company repository. Codenotary states: "Available immediately under the Apache 2.0 license, AgentMon Live gives developers access to its source code to inspect, modify, extend and contribute improvements."

The "Done Gate" check

The feature Codenotary puts at the centre of the release is called "Done Gate". Coding agents end their turn by declaring the work complete. That declaration comes from the same model that just did the work, so it is a self-assessment, not a build result. A confident summary, a clean git diff and a test file skipped because an environment variable was missing can all appear in one session.

Codenotary's announcement says the AgentMon "Done Gate" "examines recorded session activity for unresolved failures and missing or outdated verification, prompting Claude Code to finish the work or clearly explain what remains." The same release describes it as a check that looks for "evidence of unfinished work – including failed or outdated tests, incomplete tasks and subagent failures – before an agent finishes." An optional second-model advisor can separately assess whether the agent's response addresses the developer's original request.

That is narrower than a guarantee of correctness. The check works on recorded session activity. It does not make the agent smarter, and it does not replace a reviewer who understands the system. It is a gate on the evidence the session produced, not proof that the change is the right design.

Token cost of a wrong completion

Take Anthropic's published API rates as the reference point: Claude Fable 5.2 is listed at $8 per million input tokens and $40 per million output tokens, with a fast mode at $16 and $80. Token spend on a coding agent mostly comes from output. Output is what the model generates: code, diffs, test fixes, explanations.

Run the arithmetic on a hypothetical heavy day. An agent that produces 500,000 output tokens across code, tests and corrections costs $20 at the standard rate and $40 in fast mode. If the first pass reports success on a broken build and the loop has to be repeated, that is the same bill twice. Spread over one developer working twenty such days a month, the total is 10 million output tokens and $400 of output per developer per month, before a single line reaches review. Cost scales with how often the agent's completion claim is wrong.

That is arithmetic on a stated assumption, not a measurement. Our own rig has not benchmarked AgentMon Live. AI Arena tests local models, and this tool monitors a cloud coding agent.

Early evaluation, not a benchmark

Codenotary does publish one early user anecdote. During one user's initial two-day evaluation across nine projects, the "Done Gate" flagged potentially unfinished work and returned it to Claude Code 59 times. In 12 of those instances, the agent subsequently performed additional work. The examples included completing an unfinished production deployment, rebuilding software after further source-code changes and performing end-to-end verification of applications previously reported as complete although testing was not done.

Codenotary itself calls this an early individual evaluation, rather than a controlled benchmark. That distinction matters. The announcement does not publish a false-positive rate, the number that would show whether the gate saves time or adds a confirmation step. It also does not show how the tool performs across different repositories, languages or team sizes.

Codenotary's wider AgentMon platform has been cited by the company at more than five million agent interactions a day. That is a vendor figure tied to the broader AgentMon line, not a measurement of AgentMon Live, and it is not independently audited.

Local processing for European teams

Source code and agent transcripts are among the more sensitive data a European company holds. A monitoring dashboard that streams those transcripts to a vendor's cloud needs a data processing agreement, a transfer analysis and a conversation with the DPO before anyone points it at a production repository. A tool that runs on the developer's own machine avoids that workflow. Codenotary says AgentMon Live's dashboard and recorded fact-check verification, based on recorded session activity, run locally.

This is a practical difference in how fast a team can adopt it, not a compliance guarantee. One limit: this is not an AI Act obligation. Article 50 transparency rules that became binding on 2 August 2026 concern detectability and disclosure of AI-generated content, not agent logs. Agent monitoring here is an engineering and GDPR question. A deadline change for Annex III high-risk systems to December 2027 does not touch it either.

Requirements and limits

Codenotary states: "AgentMon Live is available now at no charge for Claude Code 2.1.308 and later." The release is specific to Claude Code. The announcement does not describe AgentMon Live as a multi-framework tool, and it does not list support for other coding agents. The customizable dashboard interface supports six languages, according to the same announcement.

Developers can install it from Claude Code using: /plugin install agentmon-live --marketplace codenotary/agentmon-live. Source code, installation instructions and contribution information are available on GitHub. The licence is Apache 2.0, which permits commercial use and modification as long as the licence notices stay in place.