Local-first coding-agent usage analytics for tokens, cost, timing, sessions, and tool calls.
Live demo | Downloads | Screenshot | Install | Privacy
AgentMeter is an open-source Go + Vue dashboard for understanding local coding-agent session usage. It reads local agent JSONL session files, indexes them into SQLite, and shows tokens, estimated cost, timing, session history, models, projects, cache reuse, daily/project efficiency metrics, tool-call behavior, and model health/drift signals in the private local Web dashboard, with terminal summaries over the same local data.
No proxy, no cloud service, no telemetry.
- Supported agents: Codex, Claude Code, CodeBuddy, WorkBuddy, and generic JSONL directories.
- Privacy model: session data stays on your machine in a local SQLite database; AgentMeter does not proxy traffic or upload telemetry.
- Primary views: sessions, daily usage, models, projects, cache reuse, estimated cost, model signals, tool-call analytics, and offline command/privacy audit findings.
- Interfaces: local Web dashboard by default, plus a terminal UI over the same database and query behavior.
- Release assets: cross-platform archives are published on
GitHub Releases as
AgentMeter-<platform>-<arch>builds.
Coding agents leave useful local session data behind, but raw JSONL is hard to inspect directly. AgentMeter turns that data into answers you can use:
- How many sessions did I run?
- How many tokens did they consume?
- What did those tokens roughly cost?
- Which models, days, projects, or sessions used the most?
- How stable is cache hit behavior by day, model, or project?
- Where are cost per session, cost per active hour, cache savings, retry pressure, or failure pressure changing by day or project?
- Where did model health, service drift, throughput, cache, or tool-use symptoms change in real usage?
- Which tools were called most often?
- How long did sessions and tool calls take?
- Local Web dashboard for sessions, tokens, estimated cost, daily usage, model usage, project usage, cache-hit trends by day/model/project, and tool-call analytics.
- Model Signals for raw operational signals, daily/project operational efficiency, and health/drift monitoring, including output expansion, reasoning share, cache miss, throughput, model call status when available, token/cost shape, cost per session, cost per active hour, cache savings, retry/failure pressure, low-confidence cohorts, and anomaly sessions. These are local operational proxies, not universal model benchmark scores.
- Offline audit view for command-risk and privacy/secret findings derived from indexed local session data.
- Terminal UI mode over the same database, indexing pipeline, pricing rules, and query behavior.
- Codex, Claude Code, CodeBuddy, WorkBuddy, and generic JSONL source detection.
- Multiple labeled source instances for developers running several local coding agents or several roots from the same agent family.
- Incremental SQLite indexing with source path traceability and parse status.
- Built-in pricing registry with unknown models clearly marked as
unpriced.
For most users, start with a packaged build:
- Download the matching
AgentMeter-<platform>-<arch>archive from Releases. - Unpack the archive.
- Start the local Web app.
On Windows:
.\AgentMeter.exe -startOn macOS or Linux:
./AgentMeter -startOpen the dashboard if your browser does not open automatically:
http://127.0.0.1:34115
On first launch, click Update Index in the app. AgentMeter defaults to
detected local agent homes such as ~/.codex and ~/.claude; you can add more
source roots in Settings and label them when paths alone are ambiguous. A
source instance is one local root, while the agent family (codex, claude,
and so on) controls parser behavior and family-level filters. Update Index
scans only new or changed JSONL files; Rebuild Index clears indexed files
for enabled sources and parses them all again.
For alternate install methods, CLI modes, data locations, and developer workflows, see Install and Getting Started.
AgentMeter is designed to stay local:
- Reads local session files only.
- Does not proxy model traffic.
- Does not upload session data.
- Does not require a cloud account.
- Stores normalized data in a local SQLite database.
- Audit findings may store raw local evidence so command and privacy issues can be inspected without leaving the machine.
AgentMeter is an MVP for local coding-agent JSONL usage. The Web UI is the default interface; the TUI is available as a terminal MVP over the same application core.
See Roadmap for planned work.
- Install
- Supported Agents
- Privacy
- Comparison
- Release Distribution
- Getting Started
- Project Brief
- Architecture
- UI Modes
- Data Model
- Model Signals
- Session Formats
- Pricing Sources
- Validation
- Roadmap
Issues and pull requests are welcome, especially for parser edge cases, pricing updates, packaging, and adapters for other coding agents.