Read the blog: https://clickhouse.com/blog/agentic-coding

Alexey Prompts

A viewer of Claude Code sessions with ClickHouse backend.

How to use

Create the database and tables. The raw table uses JSON type hints plus a few materialized columns and a text index so the UI's queries stay fast (see schema.sql for the full definition and the reasoning):

clickhouse-client < schema.sql

CREATE TABLE IF NOT EXISTS claude_code.classification
(
    data JSON
)
ENGINE = ReplacingMergeTree
ORDER BY data.session_id::String;

Upload your sessions:

clickhouse-local -q "
    SELECT _path AS path, json AS data
    FROM file('$HOME/.claude/projects/*/*.jsonl', 'JSONAsString')
    WHERE isValidJSON(json)
    SETTINGS input_format_allow_errors_ratio=0.1
    FORMAT Native
" | clickhouse-client --host '...' --password '...' -q "INSERT INTO claude_code.raw FORMAT Native"

This is likely optional:

clickhouse-local -q "
    SELECT _path AS path, json AS data
    FROM file('$HOME/.claude/history.jsonl', 'JSONAsString')
    WHERE isValidJSON(json)
    SETTINGS input_format_allow_errors_ratio=0.1
    FORMAT Native
" | clickhouse-client --host '...' --password '...'  -q "INSERT INTO claude_code.raw FORMAT Native"

Uploading only new data (incremental updates)

After the first upload, you don't need to re-send everything. Re-running the commands above is always safe (the raw table is a ReplacingMergeTree, so duplicates collapse), but it re-reads and re-sends every session each time.

Use upload_incremental.sh to upload only new sessions and new entries appended to existing sessions:

CH_HOST='...' CH_USER='...' CH_PASSWORD='...' ./upload_incremental.sh

On each run it uploads the union of two sets:

1. Sessions ClickHouse doesn't have yet — it asks the raw table which paths are already loaded and uploads any local .jsonl that's missing. This does not rely on file mtimes, so it works when you collect from several machines, at different times: copy the .jsonl files over (rsync, scp, …) and upload later. (Plain mtime checks miss these, because copying preserves the original, older mtime.)
2. Sessions that grew — files modified since the previous successful run on this machine (tracked by the marker file ~/.claude/.ch_last_upload) are re-uploaded so appended entries get in.

Any overlap is deduplicated by the table's ORDER BY (sessionId, timestamp, uuid), so nothing is duplicated. It's idempotent, so it's a good fit for cron, e.g. hourly:

0 * * * * CH_HOST='...' CH_PASSWORD='...' /path/to/upload_incremental.sh >> ~/.claude/ch_upload.log 2>&1

Options (environment variables):

• CH_SECURE=1 / CH_SECURE=0 — force TLS on/off (defaults to on for any non-localhost host, which ClickHouse Cloud requires).
• PROJECTS=/path/to/projects — where the .jsonl files live (default ~/.claude/projects); point this at a directory you sync other machines into.
• FORCE=1 — re-upload every local file, ignoring both the marker and what's already in ClickHouse.

Edit index.html around CH_URL, CH_USER, CH_PASSWORD with your database credentials.

Optional: ask claude to run classification scripts and fill the classification table.

Open index.html.

Do not share your sessions publicly

The sessions contain tool calls and results, including fragments of arbitrary files. They may contain sensitive info, that's why it's not recommended to share sessions publicly.