Providers & filtering

claudex indexes eleven coding agents into one database and reports across all of them. Every reporting command spans every provider by default; a shared set of filter flags narrows the view.

The providers

Provider	Source	Notes
Claude Code (claude)	~/.claude/projects/**.jsonl	Worktrees aggregate to the parent project; subagent transcripts roll up to their parent session.
OpenAI Codex (codex)	~/.codex/sessions/** and ~/.codex/archived_sessions/**	Project comes from the transcript's cwd; sessions under archived_sessions/ are flagged archived. Token usage is the last cumulative token_count (cached input → cache read).
GitHub Copilot CLI (copilot)	${CLAUDEX_COPILOT_DIR:-~/.copilot}/session-state/*/events.jsonl	Project comes from the session's cwd. Token usage is read from the final session.shutdown metrics; cost is an API-equivalent estimate computed from the rate card (Copilot bills by premium requests — the count is kept in extras).
VS Code Copilot Chat (copilot-vscode)	VS Code User/workspaceStorage/*/chatSessions/** (stable + Insiders; override with CLAUDEX_VSCODE_USER_DIR)	Indexes chat text, tools, models, and timing. VS Code stores no token counts locally, so these sessions report zero tokens and $0 cost.
Pi (pi)	~/.pi/agent/sessions/**	Multi-provider under the hood; claudex trusts Pi's own per-message cost (local Ollama models report $0).
OpenClaw (openclaw)	${OPENCLAW_STATE_DIR:-~/.openclaw}/agents/*/sessions/**	Reads classic transcripts plus trajectory sidecars. Provider-reported cost is trusted when present; trajectory-only sessions are de-duplicated with their transcript later.

| OMP (omp) | ~/.omp/agent/sessions/** | Oh My Pi: recursive session/subagent discovery, assistant usage and separate model_usage records. Honors OMP profiles and agent-directory settings. | | Gemini CLI (gemini) | ~/.gemini/tmp/**/chats/session-* | Reads JSON sessions and JSONL updates, including tools and token usage. | | OpenCode (opencode) | ~/.local/share/opencode/opencode.db or legacy storage/ | Read-only SQLite access and legacy JSON sessions. Honors XDG_DATA_HOME; includes WAL changes when checking freshness. |

| Cline (cline) | ~/.cline/data/sessions/ and editor globalStorage/saoudrizwan.claude-dev/tasks/ | SDK artifacts and legacy VS Code task history, UI-recorded usage/costs, tools, and text. | | Roo Code (roo) | Editor globalStorage/rooveterinaryinc.roo-cline/tasks/ | API conversation history, UI usage/costs, and model metadata. |

A provider is indexed only if its data directory exists, so you never need to opt in — install Codex, Copilot, Pi, or OpenClaw and claudex picks them up on the next sync.

Additional session directories

Extra roots are additive for every provider. Default discovery remains enabled. Store persistent JSON configuration at ~/.claudex/config.json (or $CLAUDEX_DIR/config.json):

{
  "providers": {
    "omp": { "session_dirs": ["/path/to/sessions", "~/other/sessions"] },
    "codex": { "session_dirs": ["/path/to/rollouts"] },
    "opencode": { "session_dirs": ["/path/to/opencode-data"] }
  }
}

Select another file with --config FILE or CLAUDEX_CONFIG. A CLI file selection wins over the environment variable. Relative directories resolve beside the configuration file; ~ expands to your home. Unknown providers and misspelled configuration keys are rejected.

For one-off additions, repeat the global flag:

claudex --session-dir omp=/path/to/sessions --session-dir pi=/other/transcripts index
claudex --config ./sources.json sessions --provider omp,opencode --json
claudex providers --json

CLI directories resolve from the current working directory and add to configured ones. providers --json includes every root_dirs entry. Changing the list invalidates the five-minute discovery window immediately. Overlapping directories and symlink aliases are deduplicated by canonical file path; symlink directory cycles are skipped. Removing a source retains its indexed history.

Extra directories contain transcripts recursively, without requiring a provider's usual parent-directory layout. Copilot CLI discovers events.jsonl; VS Code Copilot accepts .json and .jsonl sessions. OpenCode directories contain opencode.db or legacy storage/ (a direct database path is also accepted). Gemini accepts its data root or a directory containing sessions. Cline and Roo accept individual task directories, task trees, or their data parent; Cline also reads SDK session artifact trees. VS Code, Insiders, and VSCodium default locations are included. Set the provider correctly: claudex uses that provider's parser rather than guessing file formats. Extra directories apply to indexed commands; Claude --no-index continues to scan its standard transcript storage.

The index is additive

claudex's index is retentive, not a cache. When you archive or delete a session from disk, its indexed data is kept — flagged as no longer on disk, but still counted in cost, search, and history. Schema upgrades migrate in place (they never drop tables), so your usage history survives across versions.

The only command that discards retained data is claudex index --force, which wipes and rebuilds from whatever is currently on disk.

Use --on-disk-only on a report to exclude retained/archived sessions and see just what's live on disk right now.

Filtering flags

Every reporting command (sessions, cost, search, tools, models, turns, prs, files) accepts the same cross-cutting filters:

Flag	Effect
--provider <claude|codex|copilot|copilot-vscode|pi|omp|gemini|opencode|cline|roo|openclaw>	Restrict to one or more providers. Repeatable or comma-separated. Default: all.
--project <substr>	Filter by project path substring.
--model <substr>	Filter by model (e.g. opus, gpt-5).
--since <when>	Only sessions at/after this time.
--until <when>	Only sessions at/before this time.
--on-disk-only	Exclude retained sessions whose file was archived or deleted.
--json	Machine-readable output; always includes a provider key per row.

Time values

--since and --until accept three formats:

• A date — 2026-01-15 (start of day for --since, end of day for --until).
• An RFC3339 timestamp — 2026-01-15T08:30:00Z.
• A relative span — 7d, 12h, 2w, 30m (interpreted as "now minus span").

So --since 30d means the last 30 days, and --since 2026-01-01 --until 2026-01-31 covers all of January.

Examples

# How much have I spent on Codex this month?
claudex cost --provider codex --since 30d

# Compare Claude vs Codex cost, this week
claudex cost --provider claude,codex --since 7d

# What would my Copilot usage cost at API rates?
claudex cost --provider copilot --since 30d

# List recent Pi sessions
claudex sessions --provider pi --since 14d

# Search OpenClaw trajectory-backed sessions
claudex search "tool timeout" --provider openclaw --json

# GPT-5 usage across providers
claudex models --model gpt-5

# Only sessions still on disk, in one project
claudex sessions --project myrepo --on-disk-only

Provider-aware output

• In tables, a Provider column appears only when the result set spans more than one provider — a Claude-only result stays narrow.
• In --json, every row carries a "provider" key so scripts are unambiguous.

claudex cost --per-session --json | jq 'group_by(.provider) | map({provider: .[0].provider, cost: (map(.cost_usd) | add)})'

--no-index

The --no-index fallback scans Claude Code transcripts directly (no database). It honors --since/--until/--model in memory, but is Claude-only — use the default indexed path for other providers. Additional directories apply to indexed commands.

Other harnesses

Cursor, Amp, Windsurf, and Aider are not automatically indexed. Their exported, cloud-backed, optional-hook, or Markdown histories need separate format adapters; pointing their files at another provider does not make them compatible.