Agent Console

Agent Console reads Claude Code and Codex session history on your own machines and shows usage in a local browser dashboard. It's for developers and teams who want to see where their coding agents spend tokens, how well caching works, and which sessions need attention. The open source project uses the MIT license and runs on macOS, Linux and Windows without an account.

The dashboard separates cached input, cache writes, uncached input and output, with breakdowns by model, session, machine and person. It tracks usage over time and estimates cost at published model rates. Missing readings stay unknown, and estimates carry a label rather than appearing as exact bills.

Session views show activity and observed subagents. Context and cache analysis flags growing inputs and possible cache breaks; local alerts flag repeated tool calls, token spikes and spending without an observed successful tool result. These signals don't prove an agent is stuck.

A self-hosted team hub combines reporting from laptops and servers without counting the same usage twice. Prompts, replies, file paths and file contents stay on their original machines. Default reporting sends model IDs, timestamps, token counts and salted hashes; sharing project names or alert and tool-activity counts is optional. The console sends no telemetry or crash reports.

Optional integrations accept Claude Code OpenTelemetry and Kong or LiteLLM metrics, kept separate from transcript totals. Prometheus export and a Grafana dashboard support existing monitoring setups. Presenting mode substitutes stand-in names for projects, branches, machines and people.