agentabacus

Local-first analytics for AI coding agents.

AI coding agents generate session logs, but each agent stores them differently. agentabacus brings those logs into one local database so you can see:

• What did I spend?
• Which models are doing the work?
• Which projects or sessions cost the most?
• How much of the usage comes from subagents?
• Where are tokens being spent?

No server. No account. No telemetry. Your data stays on your machine.

Install

Try it without installing:

uvx agentabacus report

Or install it permanently:

pipx install agentabacus

Running it with no arguments shows the whole surface:

Then check what agentabacus can find:

agentabacus doctor

Collect your agent logs:

agentabacus collect

Collection is incremental, so you can safely run it again as new sessions are created.

To collect from one agent only, use:

agentabacus collect --claude

With no agent flag, collect reads every supported agent found on your machine.

Reports

Get a cost and usage report:

agentabacus report

Filter by time:

agentabacus report --since 30d

Group the results:

agentabacus report --by model

For example, --by thread separates your main agent from its subagents. Every grouping is listed in the command reference below.

Command reference

Every command also takes --help, e.g. agentabacus report --help.

Shared options

These mean the same thing wherever they appear.

agentabacus collect

Reads new log data into the archive. Incremental and safe to re-run — unchanged files cost one stat().

agentabacus report

Cost and token totals, with a breakdown.

--by accepts:

agentabacus report                              # last 30 days, by model
agentabacus report --since all                  # everything collected
agentabacus report --since 7d --by project      # last week, by project
agentabacus report --since all --by day         # daily series, for charting
agentabacus report --by thread                  # main loop vs subagents
agentabacus report --since 6m --main-only       # 6 months, no subagents

agentabacus top

The most expensive sessions in the window.

agentabacus cache

Cache efficiency, with the 1-hour and 5-minute write split priced separately. Cache reads bill at 0.1× input, 5-minute writes at 1.25×, 1-hour writes at 2× — on a long session this is usually where the spend is.

agentabacus tools

Tool-call volume and error rate.

agentabacus schedule

Collect automatically in the background, on an interval you choose. Uses your operating system's own scheduler.

agentabacus schedule --every 6h    # collect every 6 hours
agentabacus schedule               # is it on?
agentabacus schedule --off         # stop

Output goes to ~/.agentabacus/collect.log. Collection is incremental, so a long or short interval costs nothing and loses nothing and 6h is a good default.

This complements the Claude Code plugin rather than replacing it. The SessionEnd hook archives a session the moment it closes; the schedule also catches sessions that never exited cleanly, and every agent that has no hook at all.

agentabacus doctor

Which agents are on this machine, what has been collected, and any model in your data with no pricing row. Takes no options.

agentabacus export

Writes one file per table, plus the costed view.

agentabacus sql

Run SQL directly against the archive. Run it with no query to print the schema, every table, view, column and row count, so you never have to guess a table name.

agentabacus sql                                                  # what can I query?
agentabacus sql "SELECT model_id, count(*) FROM turns GROUP BY 1"

Main tables: turns (one row per request), turns_costed (the same plus cost_usd — use this one when you want money), sessions, tool_calls, prompts, pricing.

Supported agents

agentabacus doctor detects every agent on this list that is installed on your machine, whether or not it is supported yet:

Nothing from an unsupported agent is ever parsed, stored, or counted. An adapter is either finished and trusted or it is not there, so there is no half-supported state producing numbers you would have to decide whether to believe.

If one of these is yours, adding it is a single file.

How it works

agentabacus discovers session logs on your machine, parses them into a common schema, and stores the results in DuckDB.

Agent logs
    ↓
Discovery
    ↓
Adapters
    ↓
Normalized schema
    ↓
DuckDB
    ↓
Reports / SQL / Parquet

It also handles a few problems that can make agent usage data misleading:

• Duplicate usage records in agent transcripts
• Different cache pricing tiers
• Subagent transcripts stored separately from the main session
• Model pricing that changes over time

Data & privacy

Everything stays local.

By default, the database is stored at:

~/.agentabacus/agentabacus.duckdb

You can change the location with:

export AGENTABACUS_HOME=/path/to/data

agentabacus does not upload your prompts, responses, code, or usage data.

Prompt and response text is not stored. The database only keeps metadata such as hashes, counts, tokens, costs, and session information.

Claude Code auto-collection

Claude Code transcripts can be archived automatically when a session ends.

Install the plugin:

/plugin marketplace add tripleaceme/agentabacus
/plugin install agentabacus@agentabacus

Make sure agentabacus is available on your PATH:

pipx install agentabacus

No daemon or cron job is required.

For agents with no hook, or to catch sessions that never closed cleanly, add a schedule as well:

agentabacus schedule --every 6h

For contributors

Agent logs are not stable APIs, so adapters are designed to tolerate format changes.

A new adapter generally needs:

1. A parser
2. Discovery support
3. Registration in the adapter package

See the existing adapters for examples.

Install the development dependencies and run the tests with:

pip install -e ".[dev]"
python -m pytest

Roadmap

• More agent adapters
• Local dashboard
• Warehouse exports
• dbt integration
• GitHub Actions for usage rollups

License

MIT