The SQLite of event streaming — consumer coordination on top of JSONL files.

Brooklet adds offsets, tailing, and topic discovery to the append-only JSONL files that tools like Claude Code, structlog, and OpenTelemetry already produce. It doesn't replace your files or add a broker — it just makes them consumable as event streams.

uv add brooklet             # as a library dependency
uv tool install brooklet    # as a global CLI tool
pip install brooklet        # or with pip

import brooklet

# Open a stream directory (creates .brooklet/ metadata)
stream = brooklet.open("./my-streams")

# Register an external JSONL file as a named topic
stream.register("claude-history", path="~/.claude/history.jsonl", mode="single-file")

# Consume events with automatic offset tracking
for event in stream.consume("claude-history", group="my-app"):
    print(event["_seq"], event["_ts"], list(event.keys())[:3])

# Second run with same group — picks up only new events
for event in stream.consume("claude-history", group="my-app"):
    print("New:", event)

# Tail a file for new events (like `tail -f` but with offsets)
consumer = stream.consume("claude-history", group="watcher", follow=True)
for event in consumer:
    print(f"Live: {event['_seq']}")
    if should_stop():
        consumer.close()
        break

# Register a glob pattern — all matching files are consumed in sorted order
stream.register("sessions", path="~/.claude/projects/*/*.jsonl", mode="glob")
for event in stream.consume("sessions", group="analytics"):
    print(event["_seq"], event.get("type"))

# Glob + follow detects both new lines in existing files AND new files
for event in stream.consume("sessions", group="live", follow=True):
    print(f"New event from {event.get('sessionId', 'unknown')}")

# Consumers that transform data can produce to local topics
stream.produce("scout/stats", {"type": "session-stats", "tokens": 12345}, source="scout")

# Produced topics are auto-registered and immediately consumable
for event in stream.consume("scout/stats", group="dashboard"):
    print(event)  # Has _ts, _seq, _src envelope fields

Every event gets thin metadata auto-injected:

The _ prefix avoids collisions with any producer's payload.

• echo >> is the universal producer API — external tools write JSONL, brooklet reads it
• produce() for derived data — consumers that transform and re-emit use stream.produce()
• Consumer groups — independent offset tracking per group name
• Source registration — maps external file paths to topic names
• Byte offsets — O(1) resume, no line scanning on restart
• Path-style topics — "scout/session-stats" creates nested directories

Brooklet ships a unified CLI with core commands and plugin subcommands:

# Core commands — pipe-friendly Unix citizens
echo '{"type":"hello"}' | brooklet produce my-topic --stream-dir ./streams
brooklet consume my-topic --group reader --stream-dir ./streams | jq '.'
brooklet watch my-topic --group watcher --stream-dir ./streams    # compact line-per-event tailing for Claude Code Monitor
brooklet cat my-topic --stream-dir ./streams                      # read-only, no offset tracking
brooklet register sessions "~/.claude/projects/*/*.jsonl" --mode glob --stream-dir ./streams
brooklet topics --stream-dir ./streams --json

Set BROOKLET_DIR to avoid repeating --stream-dir:

export BROOKLET_DIR=./streams
echo '{"event":"test"}' | brooklet produce events
brooklet consume events --group reader

The watch command tails a topic and emits one short human-readable line per event, designed for Claude Code's Monitor tool which turns each stdout line into a chat notification:

brooklet watch my-topic --group watch-session --stream-dir ./streams
# #1 14:03:22 type=hello n=1
# #2 14:03:23 type=world n=2 tags=[...2 items]
# #3 14:03:24 type=result passed=True duration_s=0.42

Unlike consume --follow, which emits raw JSON envelopes, watch writes compact #seq HH:MM:SS key=val lines, scrubs control characters (a stray \n would split one event into two Monitor notifications), and reconfigures stdout to line-buffered mode so events reach the reader immediately through Monitor's pipe. It always follows — use consume --follow if you need raw JSON for a machine consumer.

For a single pytest run in one shell, a plain pytest --tb=line | grep PASS under Monitor is simpler and you should reach for that first. watch only pays its keep when one of these matters:

1. Gapless resume across restarts. tail -f gives you "last N lines then follow" out of the box, or "follow from byte 0" with tail -c +0 -f — but neither option persists a resume point across restarts, so you either miss events written during the gap or replay everything from the top. Brooklet tracks a byte offset per consumer group, so when Monitor's TaskStop kills the process or your session restarts later, watch picks up exactly where it left off — no replay, no missed events. This is the killer feature that nothing else in the JSONL-tail design space offers.

2. Cross-session event bus. One Claude Code session produces events, another session watches them. Because each consumer group has its own offset, the watcher doesn't need the producer to cooperate, and reconnecting never misses anything. No broker, no server — just appended JSONL files.

3. Aggregated streams, not raw log lines. The scout and pytest adapters transform raw session/reportlog JSONL into derived SessionStats / RunStats events. brooklet watch scout/session-stats shows the aggregation, not the firehose, which keeps Monitor's one-event-per-notification contract useful.

4. Multi-source fan-in with per-source attribution. Register several JSONL files as separate topics and the _src envelope field tells you which source each event came from when you consume across them. tail -f file1 file2 file3 interleaves lines with no attribution and no per-file offsets.

The most compelling of the four, and the easiest to demonstrate. A single shell, no Monitor required — this shows the property that makes brooklet + Monitor a legitimate pairing:

mkdir /tmp/brooklet-resume && cd /tmp/brooklet-resume

# Seed two events
echo '{"n":1,"msg":"first"}'  | brooklet produce demo
echo '{"n":2,"msg":"second"}' | brooklet produce demo

# Consume them, then stop
brooklet watch demo --group resumer &
WPID=$! ; sleep 1 ; kill $WPID ; wait
# #1 HH:MM:SS n=1 msg=first
# #2 HH:MM:SS n=2 msg=second

# Produce two more WHILE no watcher is running
echo '{"n":3,"msg":"during gap"}' | brooklet produce demo
echo '{"n":4,"msg":"still gap"}'  | brooklet produce demo

# Restart — resumes from the saved offset, only shows the new events
brooklet watch demo --group resumer &
WPID=$! ; sleep 1 ; kill $WPID ; wait
# #1 HH:MM:SS n=3 msg=during gap
# #2 HH:MM:SS n=4 msg=still gap

Same group name, different process, zero replay. The second run's #N prefix is a per-run counter that restarts at 1, but the n=3 / n=4 values in the payload show that brooklet skipped events 1–2 and delivered only what was written during the gap. tail -f cannot do this.

Brooklet uses pluggy for plugin discovery. Built-in plugins (scout, pytest) and third-party plugins use the same interface. Third-party packages register via entry points:

# In your package's pyproject.toml
[project.entry-points.brooklet]
my-plugin = "my_package:MyPlugin"

# Scan all sessions for a project
brooklet scout scan ~/.claude/projects/-Users-you-your-project/

# Current session only
brooklet scout scan ~/.claude/projects/-Users-you-your-project/ --current

# Live dashboard
brooklet scout scan ~/.claude/projects/-Users-you-your-project/ --current --follow --dashboard

# Produce stats as JSONL for downstream consumers
brooklet scout scan ~/.claude/projects/-Users-you-your-project/ --output scout/session-stats

Reports token usage, tool call frequency, model breakdown, session duration, and event counts.

Consumes pytest-reportlog JSONL output:

# Analyze a single test run
brooklet pytest scan path/to/test-results.jsonl

# Analyze multiple runs (glob mode)
brooklet pytest scan "reports/run-*.jsonl" --glob

# Produce summary stats to a brooklet topic for downstream consumers
brooklet pytest scan "reports/run-*.jsonl" --glob --output pytest/summaries

Reports pass/fail/skip/error counts, total duration, slowest 5 tests, and failure details per run.

To generate the input JSONL, install pytest-reportlog and run:

pytest --report-log=test-results.jsonl

The --output flag produces structured summaries to a brooklet topic that downstream consumers can read. See examples/ci_health_check.py for a complete example that gates CI on test health:

# Run tests → analyze → produce summaries → health check
pytest --report-log=reports/results.jsonl
brooklet pytest scan reports/results.jsonl --output pytest/summaries
python examples/ci_health_check.py reports/

The health check consumes the pytest/summaries topic and fails if any run has failures or tests exceeding a duration threshold. This pipeline runs in brooklet's own CI — see .github/workflows/test.yml.

Any Python app using structured JSON logging (via python-json-logger, structlog, or a custom formatter) produces JSONL that brooklet can consume directly — no adapter needed. See examples/jsonl_logging.py for a complete example:

# Terminal 1 — generate log events (writes app.jsonl every second)
uv run python examples/jsonl_logging.py produce /tmp/log-demo

# Terminal 2 — tail the log stream in follow mode
uv run python examples/jsonl_logging.py follow /tmp/log-demo

The example registers the log file as an app/logs topic and consumes it with offset tracking. It works with or without python-json-logger installed (falls back to a stdlib-only formatter).

# Git log as a consumable stream
git log --format='{"hash":"%h","author":"%an","date":"%aI","msg":"%s"}' -20 \
  | brooklet produce git/log --stream-dir ./demo

# Consume and transform with jq
brooklet consume git/log --group viewer --stream-dir ./demo \
  | jq -r '"\(.date[0:10]) \(.hash) \(.msg[0:60])"'

# System processes as events
ps aux | awk 'NR>1 {printf "{\"user\":\"%s\",\"pid\":%s,\"cpu\":%s}\n",$1,$2,$3}' \
  | brooklet produce system/procs --stream-dir ./demo

# Weather as a stream (via wttr.in)
curl -s "wttr.in/YourCity?format=j1" \
  | jq -c '{temp: .current_condition[0].temp_F, desc: .current_condition[0].weatherDesc[0].value}' \
  | brooklet produce weather --stream-dir ./demo --source wttr

# First consume reads everything
brooklet consume git/log --group reader --stream-dir ./demo | wc -l  # → 20

# Second consume reads nothing (already caught up)
brooklet consume git/log --group reader --stream-dir ./demo | wc -l  # → 0

# Different group = independent position
brooklet consume git/log --group other --stream-dir ./demo | wc -l   # → 20

Tip: Use jq -c when piping pretty-printed JSON into brooklet produce — brooklet reads one JSON object per line.

See CHANGELOG.md for release history. Generated with git-cliff.

uv run pytest -v          # Run all tests
uv run pytest tests/bdd/  # BDD acceptance tests
uv run ruff check .       # Lint

MIT