Data Sources

Each adapter owns discovery and source-specific parsing. After parsing, every adapter returns the same normalized UsageEvent shape, so reports and filters behave the same across tools.

Supported sources

Source	Local format	Default location	Real logs
pi	JSONL	~/.pi/agent/sessions and ~/.omp/agent/sessions	checked
codex	JSONL	~/.codex/sessions (or $CODEX_HOME/sessions)	checked
Gemini CLI	JSON	~/.gemini/tmp/*/chats	checked
Droid CLI	settings JSON	~/.factory/sessions	not yet
OpenCode	SQLite	platform data directory, including opencode-<channel>.db	checked
OpenClaw	JSONL	~/.openclaw/agents and supported legacy homes	not yet
Claude Code	JSONL	~/.claude/projects and ~/.claude/transcripts (or under $CLAUDE_CONFIG_DIR)	checked
Copilot CLI	OTEL JSONL	~/.copilot/otel	not yet
Goose	SQLite	platform Goose data directory	not yet
Amp	JSON	$XDG_DATA_HOME/amp/threads (default ~/.local/share/amp/threads)	not yet
Qwen CLI	JSONL	~/.qwen/projects	not yet
Kimi	wire JSONL	~/.kimi/sessions and ~/.kimi-code/sessions	not yet
Cline	task JSON	VS Code and Cursor global storage for saoudrizwan.claude-dev	not yet
RooCode	task JSON	VS Code and Cursor global storage for rooveterinaryinc.roo-cline	not yet
KiloCode	task JSON	VS Code and Cursor global storage for kilocode.kilo-code	not yet
Antigravity	SQLite	~/.gemini/antigravity-cli/conversations	checked
DeepSeek Harness	JSONL + zstd	$DSH_HOME/sessions or ~/.dsh/sessions	not yet

In the Real logs column, checked means the adapter was compared with real logs from that tool; not yet means it is built from the tool’s upstream code and synthetic fixtures. If a report disagrees with your tool, open an issue with the llm-usage doctor --source <source> output; leave out log contents.

SQLite-backed sources use the built-in node:sqlite module, available in every supported runtime (Node.js 22.16+ or 24+).

Check what the CLI can see

llm-usage doctor
llm-usage doctor --source codex,claude

Doctor lists the paths each source searched and whether its newest files contain readable usage, without loading pricing. Use it before adding overrides: an empty report can also result from date, provider, or model filters.

Override a location

Directory-backed sources accept a named flag and the generic --source-dir form. Repeat either to scan several directories as one source:

llm-usage monthly --codex-dir /path/to/codex/sessions
llm-usage monthly --source-dir codex=/path/to/codex/sessions
llm-usage monthly --claude-dir /work/.claude/projects --claude-dir /home/me/.claude/projects

The per-source flags (--claude-dir, --codex-dir, --opencode-db, and so on) are left out of --help to keep it short; llm-usage <command> --help-all lists them. CLAUDE_CONFIG_DIR, CODEX_HOME, and GEMINI_CLI_HOME move the default locations the same way they do for each tool, as do XDG_DATA_HOME (amp, opencode, goose), GOOSE_PATH_ROOT, DSH_HOME, and COPILOT_OTEL_FILE_EXPORTER_PATH. On Windows, %APPDATA% decides where opencode, goose, and the VS Code extensions are found. Each source page lists its own.

SQLite sources use their database flags:

llm-usage monthly --opencode-db /path/to/opencode.db
llm-usage monthly --goose-db /path/to/goose/sessions.db

Persist paths in the TOML config when you use them often:

[sourceDirs]
codex = "/path/to/codex/sessions"
claude = "/path/to/claude/projects"
opencode = "/path/to/opencode.db"

CLI flags win over config values. Configuration documents the full precedence order.

Filter by source

llm-usage monthly --source codex
llm-usage monthly --source codex,claude,opencode

The source identifies the tool that wrote the session. It does not identify the billing provider. Use --provider openai, for example, when you want events billed through OpenAI across multiple source tools.

Adapter details

The pages in this section document each source’s discovery roots, parsed record types, token semantics, repository attribution, and explicit path behavior. Those details matter when a source’s own dashboard disagrees with a report because the two products may normalize tokens differently.