Developer Tools & CLI Active Development v2.0.0 2026 AGPL-3.0-only

AgentMeter

Local-first analytics and cross-tool comparison for AI coding agents

Go CLILocal-FirstZero TelemetrySchema v3AGPL-3.0

The Problem

Why This Was Built

Developers running more than one AI coding agent cannot see how token spend, cache efficiency, latency, and session health compare across tools. The evidence sits in different local stores, in different shapes, and usually only surfaces one vendor at a time.

The Solution

The Architectural Answer

AgentMeter reads those local stores read-only, validates one schema v3 across tools, and renders per-tool reports plus like-for-like comparisons in a single static binary. It names the winner, shows the uncertainty, and marks cost and coverage caveats instead of hiding them.

Capabilities

Key Features & Architecture

Per-tool reports for Command Code, OpenCode, and Muse covering throughput, tokens, cache, reasoning, cost, tool time, and session health

Cross-tool comparison for any pair (oc-cmd, oc-muse, cmd-muse) with shared model-family pooling, output-size strata, explicit verdict confidence, and Simpson's-paradox detection

Schema v3 fail-closed report and comparison builders plus independent Python oracle audits and fixture parity checks

Read-only local readers: OpenCode SQLite, Command Code and Muse session JSONL; no prompts, assistant text, tool arguments, or file contents are read or emitted

Embedded HTML dashboards and JSON/JS data files, with --serve and --watch, plus --json modes for cmd, oc, muse, and compare

Per-tool wall-clock split into work and wait time, and per-model timed coverage that gates fixed-latency rankings

Cost labels that mark recorded spend as a provider-reported floor and cache savings as a counterfactual, not a bill reduction

Strict CLI behavior: date-window validation, fail-closed empty sources, actionable wrong-side report errors, ANSI-aware width adaptation, and documented exit codes 0/1/2

Six cross-platform release archives for Linux, macOS, and Windows with checksums and AGPL, NOTICE, and third-party license texts

Engineering

Technical Specifications

Primary Languages: GoSQLitePython
Frameworks & Tooling: Go standard librarymodernc.org/sqliteEmbedded HTML/CSS/JavaScript
Architecture & Execution Model: Single static Go binary with read-only SQLite and JSONL readers, schema v3 fail-closed report and comparison builders, embedded HTML dashboards, and six cross-platform release archives
Licensing Model: AGPL-3.0-only (copyleft)
Release & Lineage: Active release cycle since 2026 · Maintained by @JadMadi

Knowledge

Frequently Asked Questions

Does AgentMeter send my session data to the cloud?

No. AgentMeter is local-first and zero-telemetry. It reads the local stores already on your disk, opens the OpenCode database read-only, and makes no network requests.

Which tools can I analyze and compare?

Command Code, OpenCode, and Muse. Any two can be compared (`agentmeter compare`), and each tool also has its own report command (`agentmeter cmd`, `agentmeter oc`, `agentmeter muse`).

How do I install AgentMeter?

Use a precompiled release archive from GitHub, or run `go install github.com/jadmadi/agentmeter/cmd/agentmeter@latest` with Go 1.26 or newer. `./build.sh --all` builds the same six archives locally.

Why is recorded cost called a floor?

Recorded cost comes from what the provider reports. Free, contributor, and subscription-tier models may report $0, so per-step and per-1M figures can understate paid usage. Cache savings are a counterfactual at fitted prices, not money removed from a bill.

Does AgentMeter read my prompts or code?

No. It reads usage, timing, model, and tool metadata. It does not read or emit prompts, assistant text, tool arguments, or file contents. Session titles and project names are included by default and can be suppressed with `--no-titles` and `--full-paths`.