JSON output

Every read command supports --json, emitting a pretty-printed, stable shape. This is the recommended contract for automation — it doesn't change between patch releases.

Examples

# Top projects by cost
claudex cost --json --limit 5

# All sessions as JSON
claudex sessions --json

# Summary as JSON
claudex summary --json

Piping to jq

# Total cost across everything
claudex summary --json | jq '.total_cost_usd'

# Project names only, sorted
claudex sessions --json | jq -r '.[].project' | sort -u

# p95 turn duration for one project
claudex turns --project claudex --json | jq '.[0].p95_duration_ms'

# Every PR this month
claudex prs --json \
  | jq --arg m "$(date +%Y-%m)" '.[] | select(.timestamp | startswith($m))'

Shapes, by command

Per-command pages have the authoritative shape; this section is a cheat sheet.

summary

Single object. Keys: total_sessions, sessions_today, sessions_this_week, total_cost_usd, cost_this_week_usd, total_input_tokens, total_output_tokens, total_cache_creation_tokens, total_cache_read_tokens, total_tokens, thinking_block_count, avg_turn_duration_ms, pr_count, files_modified_count, top_projects, top_tools, top_stop_reasons, model_distribution, most_recent. See summary for the full shape.

sessions

Array. Each entry: project, session_id, file_path, date, duration_ms, message_count, model.

session

Single object. Keys include project, file_path, session_id, date, last_activity, token totals, cost_usd, turn_stats, models, tools, files_modified, pr_links, stop_reasons, attachments, permission_changes.

cost (aggregated) / cost --per-session

• Aggregated: project, sessions, input_tokens, output_tokens, cache_creation_tokens, cache_read_tokens, avg_cost_per_session_usd, models (array of families), cost_usd.
• Per-session: project, session_id, date, model, models, input_tokens, output_tokens, cache_creation_tokens, cache_read_tokens, cost_usd.

search

Array. Each entry: project, session_id, message_timestamp, message_type, snippet, rank.

tools (aggregated) / tools --per-session

• Aggregated: tool, count.
• Per-session: project, session_id, date, tools (object — {name: count}).

models

Array. Each entry: model, model_family, session_count, input_tokens, output_tokens, cache_creation_tokens, cache_read_tokens, avg_cost_per_session_usd, avg_tokens_per_session, service_tiers, inference_geos, avg_speed, total_iterations, cost_usd.

turns

Array. Each entry: project, turn_count, avg_duration_ms, p50_duration_ms, p95_duration_ms, max_duration_ms.

prs

Array. Each entry: project, session_id, timestamp, pr_number, pr_repository, pr_url.

files

Array. Each entry: file_path, modification_count, distinct_session_count, last_touched_at, top_project.

Why not CSV?

JSON round-trips nested structures (token breakdowns, model distributions) without flattening. For a CSV equivalent, jq -r has you covered:

claudex cost --json \
  | jq -r '.[] | [.project, .cost_usd] | @csv'

Stability

The JSON shape is the public contract. Fields may be added; existing fields aren't removed or renamed without a major-version bump. Field order is preserved within objects because serde_json uses insertion order.

If you need richer access, see the index schema — you can query the SQLite database directly. But JSON output is the stable surface.